Merken

# Video Similarity (30.06.2011)

#### Automatisierte Medienanalyse

## Diese automatischen Videoanalysen setzt das TIB|AV-Portal ein:

**Szenenerkennung**—

**Shot Boundary Detection**segmentiert das Video anhand von Bildmerkmalen. Ein daraus erzeugtes visuelles Inhaltsverzeichnis gibt einen schnellen Überblick über den Inhalt des Videos und bietet einen zielgenauen Zugriff.

**Texterkennung**–

**Intelligent Character Recognition**erfasst, indexiert und macht geschriebene Sprache (zum Beispiel Text auf Folien) durchsuchbar.

**Spracherkennung**–

**Speech to Text**notiert die gesprochene Sprache im Video in Form eines Transkripts, das durchsuchbar ist.

**Bilderkennung**–

**Visual Concept Detection**indexiert das Bewegtbild mit fachspezifischen und fächerübergreifenden visuellen Konzepten (zum Beispiel Landschaft, Fassadendetail, technische Zeichnung, Computeranimation oder Vorlesung).

**Verschlagwortung**–

**Named Entity Recognition**beschreibt die einzelnen Videosegmente mit semantisch verknüpften Sachbegriffen. Synonyme oder Unterbegriffe von eingegebenen Suchbegriffen können dadurch automatisch mitgesucht werden, was die Treffermenge erweitert.

Erkannte Entitäten

Sprachtranskript

00:01

fellow at my pleasure to welcome you to the lecture multimedia databases today and they will be going to the deeper into into videoed trieval on last time we

00:17

were a kind of pre prose testing the video we were on Abstracting the video and segmenting it in the end it in the shops Austria's remove images of images of some of the beauty order and now examples I'm of well and use programme on TV a way of the 2 pickled elements like the and command like rip-offs from some locations bomb people are giving opinions on something on describing something and that or a caught of something like that and and is all different the from shops that could be represented by a single frame of making the video little but I'm easy attitude to manage because of the beauty of this is the very storage intensive for and if you do shoppe detection take just want he frame of each of the videos that's fact to something that is far less than the actual but you don't have to handle the video on the Web talking about some some some model of salt to shoppe detection from starting from from simple wrestling model suede could just that something like OK if it the colours changed to much land justice units in new shoppe to statistical model of Load look at the structure of the motion Bektas and said Well between shops the motion which was a bit different from within shops and so we come from some of ways to exploit distribution statistical distributions to find out which of the and the after you the compost the video into a set of key Frank into the video attraction for do we do now well would want to do now is kind of compel videos of a

02:27

symbol of all not find a measure already this these videos of the week and and find out what it was all about and what sort of a stop it is to be a celebrity a would talk about basically 2 ways of defining video Summerteeth the on intuitive of the of the ideal videos the and the of or not to the of bomb

02:57

determining the simulated between video is also close of area MP or for the ranking so if I'd want a typical period island that video with the Preston in that it would have found that and we have to rank all the videos calling to halt similar they onto the video that we managed and how we do that should either by mad that at all but just providing the frame of the of the with looking for that is that something that will talk about but it was also interesting for finding duplicates sell for example if you have a on a two but the video Paul like like huge you what would you think is the number of completed on duplicates such a public what's considerable and storing all these do but it is not a very sensible thing to do is it so if you find duplicate can that already in the bag but we don't need to stalled and another applications that are immediately springs to mind is done of is detecting copyrightinfringement if you know which videos are similar to use of the deal with the of number the music industry because many people in the community just up loads of music videos to to you to open to might of the deal 1 of the name of awful and of false that is not their content and not the continent of cell finding the infringement also also need a description of Hollis similar of the of bomb and was Civiletti we of costly to mesh the some some degree of seminary the that we can think of it this treaty was more lessons to some some some and of course if we see it in the evening I began to say about just take the famous old even only keyframe and the celebrity between videos is the percentage of traces with a high Brazil symplectic and we already know how to determine high visual celebrity because we can do that was in no mood unit which we at the start of the colonies to brands symbol of the state of some of the same shapes of role at enough we we a you were through the and them this kind of similar to stun in intake space with 3 other tiny model celebrity way just have the tough talk that efficient and say OK take all the words and take the be all that between the words and and and of the by the by that the number of full of operating and 9 0 house and the stakes are high sell this kind of what is often referred to as Tanimoto celebrity in and and take a wide a we just that the beauty of its the straight for the idea

06:19

of and why renowned need if we want to do it is that we need to the identification of result features from the fact and the interesting part is here that we do have time series of the chat with just don't have the 1 feature but for each image with each framed we have which features of that means we have a 2nd of the 2 victory the future of change over time and again we can take a couple distribution but also take something that we didn't have been images and still images of the for example motion Bektas began really term motion victims from from image to image and that is something that can help us out and for Efficiency reasons we probably should not determine the celebrity between between single frames but but rather between stops so we take the On the keyframe all of the shot and just compel keyframe with suspected of and each keyframe represent the salt and the famous symbol of the shoppe must be similar and we can go from that is all way of of of defining us in

07:39

the air to soul a few consideration of that we do have to make is that if where my number of feature the bed of a simpler to measure and were because we don't only for what colour but also textural shapes motion victory was 1 of more but of calls having more feats of also means we have to do more calculations with told of the with to do more measured and given that the video consists of multiple shops of multiple Kiefer for this multiplied by the number of so the bigger the feats and the most shops and in the video the bigger the reputation in terms of some levity for the video becomes and that this is fine wine vinegar retriever usually you would you would rather see what at IEEE sacrifice some of the acuity are too to be efficient to be more fish efficiency as the of the crucial part of the tree and think about what looks like like who will be of all you need to know if they are going into celebrities like was the image such at least to some degree recently lembit the time that would have to do CompaorÃ loads of the of and the efficiency of the of

09:14

the the interesting thing though this brings us taught 1st the 2 0 against each other how big the problem actually is in the case of the queue ok so on is the Oregon mentioned you to is a baby but would be the basis about 65 65 thousand of the day was uploaded each the and 42 40 for this kind of 1 a month ago there is a high but will be a bit of these views are up just imagine Ivanova new a new movie coming out the lot of the guys will upload trailer order to a new poll about what people would up more began a presentation of the of the boss and the bags of what went on 1 side for you to they would really like to eliminate bills will be paid 10 those for West data and on the Web either side of the world over for use if I'm going to search for something and then I'm going to go to get a based with of all the same you may be with different inquiries maybe we owe small difference in the 5 boroughs log on or maybe with a few frames more or less but the sea and though it would be actually a great idea if you were able pull identified those will be getting last attempt to get the answers to these out all the egos but that kind of stimulus with respect for your be these are not the last and so so for example if you have searched for the landscapes tonight for some reason my book or so that there are multiple duplicates for 4 days the duel here in the USA is based on a new book of the year of the use 1 is 1 of the last 1 the in of and I'd get the 1st pages with actually to people that were there being a great player and the 1st 1 with by on all print the about the gates the 2nd 1 with but not in the sense wants widened to was and concentrate on what I'm interest but at the end more spin off and questions here is what that means that whole family defined and now

11:36

it would be easy so 1 could say that the gates of the duo's which air exactly the same thing so the same free number within the but the and most of the cases we don't get the same yield case we usually the of new so of this is that they would in a introduced into thousand 6 and he said Well near that gates of about the same age as the put in at either the file for my for Beijing quoting found that more some by the since like the brightness or lightness of the Bay of maybe some points on a baking will Poisson's for example you it's quite for Monday when someone out of the new group will be brand that we do it with 1st by boat something like this high uprooted or the and the and the and utilities to introducing was suffering that they may have a different way

12:37

again the landscapes tonight example of here heavy an example of the of the 1st 5 frames of these legal the same stuffed with some eastern France only this 1 was about this 1 was about but it's still the same the your here but more frames in case the another result you should maybe with some banks introduced and so on so you can have the front modification for example here scaling here it's a bit of whiteness modification under the it would be great white painted by such cases below them together and across and so

13:24

that you can get the whole thing is just over an hour or so much eroded mountains various on the Web just aren't they study for new book the landscapes denied the players undoable 100 out of which for people per cent at the end of disease huge and if you look for the good of out of egos again with huge percentage in the so just group this would help log in easy about and the user when a pawn in searches but

14:03

OK now but we got the about how biggest public but see how we can we can solve this and summer a up by the idea was found that we have to compared the individual frames and seeds other similar and other simpler means basically computer feature compel feature but I'm off costs blotchy to make up for every official over frame work so we need to restrict ourselves to do justice more of justice more number of teachers are that we are the fruit of a representative media with minimal error and that is very the dependent on the time of the deal that he actually work on this is very very interesting time to look at collection and and see what when he actually I'm in means that you minimize the distance between the media and the representations and for example you can just take up of the tomatoes in N-dimensional space in now and then for example the method of the squalor came means with to to classes what of 1 of the interest and this is the idea of building a are of the deal signature sold for every video you get the number of seats of actress that represents a video with respect to all the other videos in the code which was minimize the see that the cost well chosen and Mike frustrated find the reputation of fine if ICI but the classes are kind of very fuzzy of Distributed obviously need some more information to distinguish between the distinguish between Clusters and and how would add some more features making the problem Hobba because my feet become bigger but making the discrimination between videos that this is the idea off of the video

16:24

signature bomb so then we just too soon for for the rest of the lecture at each frame represented by by some feature victory in a match explains and we have the measure the attached to the end them and you tedium space and just the beauty and distance Allwood of which saw although of all the of methods and all the other was that will see them in the in the lecture basically can apply to all kinds of members kinds of the board of huge tools and has said that said very deep and what you choose their depend on new collection and a boost for the basic who want to do so I'm not the celebrity picture measure for the deals is obviously Invariant with respect to the for the 2nd time so if I'd just have video containing of to shops and the swept up in the same video or not while in terms of their to is obviously the same because they don't have any other any other frames in terms of the understandability of the of of the of the of movie of the order of maybe it might become hot on the standard because you enough flight disintegrating the storey lack of life that new rather Thorogood on the technical for some of the technical notional of them their teeth because I'm otherwise but it would be very simple if we have to take the issue of seconds into account it would be very simple to trick all algorithms of also Murphy a measure into by assuming something that's not the same that it consists of the stage of some of wrapped in and and some shops that don't belong to the video or something like that immediately the I'm the sum of because so we will not look at the at the shops equities we would just see the sex of future that UK and and that's all you feel on

18:54

well now we can say that about the way some distance measures which is basically dissimilatory see between 2 feature and us all represented by the same by the by the frames and I've busily similar if US and there seem measure is beyond some so we allow some at Salam of mistake but that it and its more than Apslawn lifted levity is higher than that of long YWU just say this is not simply a NEMO and will not count as a symbol of the husband and very very often actually so the approaches Quach rep to to most people do it stood in the way

19:50

want to do now is to want to compute going to Tanimoto measure the percentage of similar frames in the video and if we do it in and nice EEF weight and we just say OK this the total number of frames of the video P that are similar to Adeste 1 frame in the other video invited by the total number of UK and for every frame in the 1st videos and look whether there was a corresponding frame in the 2nd city of them in the 2nd bed media for every frame look it there is a corresponding frame in the 1st 3rd of and and I'd invited by the total number of trades in both videos and that it this number and this

20:46

was caught my the self for help the frames and in video X I'd look whether there is a framed in video why such the distances smaller was correct restrict function being 1 of if there is being 0 with a smile you estimate of the same duo frames of why directory the function of the fact that the smaller the of and then I'd Evite by the number of friends and this in the East video symplectic the and them the videos and that is what you if each frame and ex commitment to some some of them and why in the frame and white command to from some of the acts Villa somebody 0 if there are no similar claims to video so me the frame and and a new video and map of the UK and seems that good idea than the yes or a up at that is not because consider the case where we have to images in all video X that his eye would just abbreviated the future Vectra's of care and and I've my video why but it but of the BP of camp iles similar on offer that probably not because he is a lot of content that is not in excess of Kent but it does come to the video X perfect automatic to the same job the and of course they have to 0 2 0 my account for the from the length of the video it's not it's not clear that there is correct matching so each frame of excess match on to applauding frame of off wide and the same number of stops and income from so that it is not that they have to do it and and on both sides the otherwise you will not detected dissimilarities time it sold good idea but it got my east of car

23:30

public East for for a reason and in the East stemmed from not being into it because the same take that I'd just it was the videos and be done and also by by duplication of the so for example if you have a shot that Kurds in a similar fashion in a different way the M by taking the 2nd video and a duplicate to stop but off times but take 1 child from video 8 of the excellence of but both were from video X of the and the night generated with your wife but just duplicating the French of the vote the and the might video summit between the 2 videos get closer to 1 the moral from a duplicate this show this symbol shot the zest is accounted for different shops in either of the video so it does not

24:50

work so it left the forget at the and example I'd take 1 video X over here and that has to stop and this is my feature a friend in the space of Kent where that it and they take a 2nd video this is why and these are the frames of what some or all Distributed or what is the videos and well he would say Well obviously a full full for the video excise to have this sort down here and this sort down here and witty Wyoming as this sort of out the OK surge should be around 50 per cent if you calculated in the need for video and their have a lot of funds and saying this is a member of the system of the system of the system of the system of this system is that most and I only have 1 of possibilities think of it this is a December so the book what would it actually Calculate is the biggest America of mind he said it's not impressed with his wife quote might have something to do with distributing the frames of the video all the features if rather than just measuring the distance between individual fit a trains and and say it's a similar all not and

26:35

this growing getting at so we should we should consider the quantities of similar frames as fundamental unit which look at the last 4 frames rather than on the pitch at the individual friends and if we do if we said we we get the tempo structure so we we we just see the set of which we compiling the vigil of is also similar which are we combined the vigilance similar features to classes and then account weather over the last are made up by frames of both videos because if there is a cost that is just populated by freedom from 1 video and there is a big this summer to and the degree of this summer to the degree of over that is basically the percentage of shed last rather than the percentage of shed frames popular yep well now to face but on to the same class the if the distances beyond some at the proamniotic has how do we cluster what we do and that there are several possibilities best completely Clustering with singling clefted as a lot of different costa-gavras and the big problem is consistency discounted up with a couple frames and where frame accent a frame wines from the distance of smaller at the event and we have from the frame why and that and that is the smaller 2 1 about the transit if just the distance between acts and it was say something about way the week Holland because if we consider the feature space it might be the case that we have and why that and why these allways may be smaller than exit we can say that this is falling the other 2 a Kent smaller to Epsom of the Blue Nile authors of could be the case that's like that well clear of the eggs and why and then do and this is more than happy to go on this is 1 of the and bases also smaller of firm could happen but does is no basically no guarantee for that and a lot of the big problems with singling

29:51

last it in a single enclosed straight that distance smaller than Apslawn between 2 members of the cast means they belong to the same fast but not lycée where as if to object or if 2 frames to the same class that it does not have to mean the distances are all the more because it could be a good evening be kind of like we thought was the extra money comes why this moment to long and and comes led from the Commons a M from the Yemen of the and we of such a change of all 4 frames allways with small over at the but this may be out for the stop classes can be generated for use singling clusters of and from their not nice wrong Clusters like we would like them but they could be chains basically spending the whole of some the of the whole feature space as of is not very helpful from year back so of what we do prop that's just interviews limitation on for some video X we take the Quebecor with law those cluster of the things the frames in excess that all was in an excellent sum from and we will call at last Exelon compact is full of praise of across the have Murrow's the distant of the suspect which of so if the case is like that perfect close ups of some compact from basically the distance between all 3 of them smaller that if the case is like that this is not compact because the distance for the year may be out UK and the following flotation and if we have a poem the reunion of the union of the 2 last this book you would just do the same thing so we do not yet whether the frame is off on acts of why you just say that Apslawn Cluster for both kind of Frank if and its excellent compact if they are most of features at most and at from each other yes yes that's right different

33:14

shocked who while lets don't think about fading so much thing about a at to be to use the programme from the for example you would be and command that is returning periodically are doing 1 of a report Alten done comes some Island off the beaten polyamines told images of some politicians more whatever now and then its again the and command it's the new shalt you may have moved to man the different background image from this slightly different think of the new breed helps may be Ademole warranty something like that I'm comes again Gambia income and new background him so again slightly different and soup with different shops may definitely move over time they will be some all similar exactly but the from shoppe to shoppe that may move so full fall for many types of video we will have a compact at long last the if you consider that this is the kind of news anchor the of the new think I'm and you have the Tobago image earlier and the spot where change and the rest of the image will probably be the same and you will move this cattle whenever I know but that's not the sort different types of because of what it may be that you are part of the match but that is different and the rest of the match says say so in terms of the future but that will be a typical cases where different or Curtis's different shot of of of the and command will result in different points that off Hall and at all be excellent compact from but if you have the different of the images in between 4 of you may have though of the well in the Rock which is kind of like the that landscape and now something like that and then you may have and a totally different images of a politician or something like that and and that could well be that the point stop moving that now why they may be similar but that they start moving a part and and this would be also last because the kind of the difference between either 1 of the misses of the maximum no after a compact and this is still the exit of the so this is just video exit from the programme for example of the way if I'm not have video the next day his new scuffed which is my video life than up all we will have a similar fate is over the top that will give me of some ecstasy and a obtain different images and background the same and command Saints and studio a and this would be X and why the cost at the of the bike and Ikechukwu just you look the following the while the Pope all we also contains some or all of them might be some here in all might be new classes that come was X might be existing clusters the with the events that led assuming that 1 1 day there is so way on the rock of the big of the big issues of the day will have a loss of politicians to know and and doing things may be different politicians that might be the last down here a that some of snake OMX and of course to of the reports from from year and and that will be the landscape from the landscape of the soul of the like the cost of of the next day there will be a big problem with the order of spill in the Gulf of Mexico off again in the end command sure with that some for this Clusters there will be politicians saying something about there will probably at this Costa that there would probably no Puno does emerges from this cost of will not contain any white videos but there will be some some some images from from from either now like like the beaches of the sea with the order bills of the birds that that have boiled entrusted feathers of something like that that will be a new class of some of every last has to contains range of shops of those but it but there may be 2 people elements that are very close to each other and both videos the may be elements of totally different and this whole wheat how we distinguish enough I'm so if we say we are basically its at them that the new programme that only deals with political issues are than they will all have been a part of this club said all off of discussed and so their very simple in what they do I'm you we can see that music programs from different days may be the December because they just chalet in command and public that it would sell the visual impression of the of the shops in UK and the 1st question it but find

39:36

itself week and we find the ideal videos and their to as the percentage of last taking video from those taking the shot of the day from both videos which contains range from those by resited to the total number of just of the more Clusters that we have the containing excesses and but and less Clusters we had containing only access all containing only got the more similar on the muni by taking across those follow the features and for each faster you I'd take a characteristic functions the set up a are images from the X and the cost and and the director of the fund is 1 of the last year and are images from wide in the 1st progressed from the line and is you so this told the the evaluate to whom if effect features of their frames from those videos in a otherwise 0 at a summit of anti wide by the number of total number of classes of his that should do the trick 4 but

41:26

showed I've to videos but too frightened to shops and that he for 1 is the video that as shown by the excess over here the other 1 is a video of that is shown by the thought of him from a classes to IAAF have 1 Costa because they have different of an at 2nd last and the wrong way and the cost of self called the is the ideal videos and levity well it's a further because of the 3 Custos only 1 the contains shot from both the US and a don't care if there is a 2nd across board their multiple points here from the doesn't influence the result just pounding class does not counting individual frames and mouth so trick with duplicating 1 framed from from the other video and and pudding it into the video doesn't work and UK and yes 1 BP has set off a bit it the that the the book all by and I'm not sure if it's the half would be into because the of free kind of of frames and I'm the Sri frames if we could fit it like that of the early Tony different for this might be the desert see this might be and was still see and this might be 1 of the and command order of so that the with a guest you could all you for its and they were in the end man the same and the and and half of the continent this some all different and it should be a half or you could say found well basically I'd just on the different expressions of in terms of overall frames and care from which revealed of actually are and there are 3 kinds of impressions here the Ausbil the desert the and command and and all the OS just coincide on 1 of them which makes it metal taste for with big usually have defined like that but you have to take the of the of the ground troops for for all the different images of interest from the likes and assumed to find the video is missing something because I'm you also have to cater for the case of the videos and not have the same number of talks so that might be some plastic via and some Clostridia was a summer to between the videos Mao is still a one off because 1 video as half of the of the law should be smaller Agassi to become smaller of mind and so far only taking the total number of clusters of good idea Adam but Agassi could or for for both things but so that the basic ideal

45:26

of ideal I'm video Summerteeth and now if we if we have to calculate that the mean seems to think rather simple but if we have to calculate that we have to calculate the distances between a lot of players because they see the we have to make up some the cost for the different trains and and look in every clusters whether there is a bomb a welcome there is representing frame from from from each video that kind of part of the time consuming method so we could do is we could state well maybe we don't want to look into all over classes them although the heads of all 4 frames but we would rather like a sampling approach to say OK we with which randomly take some classes all of the foetus based look into them and the bomb that the number of of dustless that contains some frame from both videos is kind of a representative for the total number of justice contains on so far of future space number of those in the future space and say Well to this whom this from this from UK and this is a random sampled related band called the if my sampling sizes but now it will be good at approximately should not of what happens in end and the cost of UK and I'm UK at the and the idea would be to represent which video through M randomly selected video frames and estimated that the idea of video severity by the number of a symbol of heads in the sand the UK and whom the problem is that

47:39

the work because it might take summer all values of of 4 4 4 am to speed up the calculation for small random samples sympathy and to something it may seem the early this told the results because it Beckham said some videos of the same length and for each frame and acts that exactly 1 similar frame and why and what tubers so that it's really the case that these are the same videos then the expected value of similar Paris with respect to M on foot of draw a frame and draw frame from the of the video the probability that I'd exactly drew the match in Paris is on but for a really know because the via a sample size and them and and the video that its could erratic in the sample size devised by the possibilities that effort to draw from each image which is the same length as the that the number of times act of number from the line up in last it takes on average the squalor route of samples to find at the East London similar the squalor route is not a very good around them some because of the squad who are already on that kind of give me at least a similar up at and we were assuming here that they were only matching pay for this is the very bad idea to do with West Ham the figure of something up and we don't want to compute all the Clusters and Competa all the Clusters we complot with random samples because that would result in a very bad this for for for we see what we do well the idea is to be buying it up the features space that that's not look at the cost of this look at the future space and at politicians of the features and politicians of all the features space directly carry the notion of a bomb or

50:22

diagrams and this brings us to on extinct took up about their job I got my looking so we want to audition a space and a field of mathematics this with his destinations and should be will not be the beginning of this was Russian mathematician or green and Warrington and his diagrams a well known for its so what actually has done he has become pause limited spacing distant by by using some sort of the world when he was with of for the sort of public was starting from a a big space and the step up 42 points from this match explains the Bulls would be by the spacing exactly the kind enough to say once all the number of 42 points by and so that you need your despite its there is just 1 point from X so it by what they here and like look for here than out there this possibilities of this was to be brought under the your of possibilities for about what they should be space and as I've said these 2 regions on pain in each of them just 1 point and the

51:48

ideal for on in his 3rd dissertation how what was that if we take the same space and again the same finds and split and some hope in the sea regions for example in the 1st regions every other point state is point she of these closer so for that would Indian Point in death regions result cost would be so based distance yet use more land based systems he in this way you can open and antique destination of this you should take a more complicated the situation with more than 40 dewpoint your been something like this and again the fault but also for example the and and were to do just that and 1 point on which he based point you is closer to the point and one off this week and the and the 1 August 1 for his wife so at the scene vocal but the but they useful and we should take the open space 42 pay out of the points they knew have hyperplane between the for points cut for example in order to pay for it to 42 points between the 2 there is this type of plane computed adult of points which are equally these the and for both of them both of these audit and and the public these days of the left of this type of range have only points which are closer to the West Oregon so this 1 here and with the right a heavily points which are calls for this 1 and probably should think of the world would be and the

53:45

hyperplane you can or a team mate in a possibilities to a computer and populated by the 10 example she so I'm going to divide the space but party should be space based on for point and points timebomb to copy of whom region Plavix 1 index and so on bed this point for this point here there is a hyperplanes which goes exactly through the middle of the distance land between these 2 points and 4 predicted that best hyperplane and each point on the high proclaimed is equally be simple bought of this 1 what is basically a sense is that the points for the Red everything was quite launched for my Repligen point and everything here last so I've or a bit of being a part vision for these before points locate non going for for foreign based again no 1 these before young saw the sick and Bank of this 1 and were to do it again I'm Computing might hyperplanes and night can also so I'm or a despite here from the previous comparison thousand below for my regional regions and they also get this is how it was before that these 2 0 despite yet again thousand on the might of whom are region for this point here and then I'd go for the with that but went into exactly the same thing again this hyperplanes here this was pumbaa 2nd point these 1 of these was here from the 1st 1 and then the views from the top down and this way a give my final 1 region when basepoint yet and then I'd the same for the 2nd points where there has been point out that in the world region point that 1 would have won for the 4th time this idea that the fine of 1 might the same

56:08

age and the as you can see where this is simply stand out of way it has of complexity offer on whether the and by half a compared 42 regions with all of the points but that is simply state of with my minority able to compute the reception and though debt at the end of the application like the advanced 1 0 which is where the money is for the growth of Greece boats your called Renault is he the 94 cents up like the when growing Chris those used up with some regional points like in the world of case and in conditions of the same of the temperature of the Crystal Skull and need somewhere where hyperplanes and you can now buy it is based on the former the and about their also somewhat efficient ways of making these issues results and complexity of and log on all over the world on my back

57:12

Oh but that was then and this with with their reasons so of for that we will continue by storm you how to use on my background for your simulate sell the various Emirates eaten for my difference with seen that we can computer for my grammes quite quite efficiently so that a good idea found and basically the were Noibi at diagrams nothing but divisions of spaces if we take the fastest As the smallest part from of the videos and the team and we can division of politicians the space the feature springs recording a plastic each Costa the of long noise regions around it the Lions and the further public last those the bigger will be the or my each clusters of their close to each other that will be very many small for more than the 5 video but has a number of frames the Vaughan audiogram for this video is a division of the future space in basically L what happened is a take the features I'd taken video X and output video X into the future of the resulting in its frames X 1 9 2 3 4 6 5 and so on I'm to expel being part of the future than my they are paid and the sea at the tessellations like that like that but that what branch of the year of year of the of the time and only this would be of or Noyac tessellations of the space for the video X Men and that by every video of that the law might sell

59:35

for some of frame in the City contains all the victims which like close to that frame band to or the other frames in the video from so about this for my diagram list some of the disc and this is basically some by the frame 1 or them only see to evict lying around here somewhere in the south UK closer in distance to X 1 pm and 2 am I like to fix re also on baby and some Hall to this fact and were will say that are a before Loiselle of some frame is basically the feature victory OS for which the palm for which the closest just say this means the closest with respect to exit is frame extinct and Frederick the of causes of from all the book at so we will see the S just a a distant function telling us which is the closest frame of the video biography to the end but we have to I'm he also set of several frames I'm you you just take the frame that his next to some some some predetermined point so the might of cost beat point here so whether they belong to a fixed 3 0 8 5 bus ride on the on on the middle firm that it has the cladistic though just say Well its equal distance we will always with a tribute to a point that is closest to the origins so we would decide for 1 direct and then sale for 1 thinks they point in phase and the and the say the of more points out assigned to to 1 point of their sold the point over here and the point of it all the time to stand up and say if are we not talking about the classes yet but talking about the frames of videos bomb again a number of that this is 1 origin point for way of the future space that these are the frames all key frames of the video UK and and and the talks between the are basically all the feature because the around it from possible future may not be justified by any framed in the video this is a possibility that there might be a frame that slightly and their closest to 2 1 of phrase the idea as to why we are doing in that way the idea of cost that if any of these feature points would be instantiated by some of the video while 1 a k and they belong to the same from this is that is the idea weakening the Vaughan Oiticica lation on the future space for every individual video of Kent and weekend uniquely identify lust signed any point in space to 1 of the frame in the video except for all points he on the board of Alliance exactly equidistance was suspected to right and that we have to decide for either 1 but it consistent fashion went so we will always be side for the one framed assigning to the 1 thing that is closest to the origin of all fall features that if UK and this doesn't the influence and this is just a minor number of points to be being on the on the board alliance between whom ourselves and which deciding for you to 1 of good OK so the

1:04:36

modernised also are combined for frames of identical cost us so we can say that if we have a point of across the board the law noise sell of the pasta is just the you need the Union of Liberal noise claims that the frames of the features the future victory was out of the frame is being part of the cost a so if I've had a couple of races and they say this is basically a last because this is more than happy to go and this is more of them to the UK and this is all the more of the for mclouth over noise cells overnight sales of the cost is basically for that the if the union of the individual who are myself 1 said that FIA good then

1:05:44

I'd 25 in similar out or not region for to videos now we have to videos and that 2 were respective while annoyed by grants from such the we saved well we have to videos we have given Epsilon was basically the at MIT a book this at is a possible the then we will look at the long Noisette also of the frames acts and the wrong noise sales of all the time why and that in the spectrum with respect for all the payers would have smaller distance off at wrong for all the ex wife payers below to the same Kaka if to range from different media UK close to each other and also them or noise cells with interest but put why have the the simple case here the raw myself for the exit and then I'd have a video of the year and year that will make the lawn myself for the Blue video locate by now say these are very close the law noise cells in the Sept UK and South if to frame from different videos close to each other but they were not myself with the SEC and the more sum up as the more intersections eye would have passed the such area the volume with the lucky

1:07:40

but that's a bit too videos by look at the video he and this is the ex video and are looked at a wide video having 2 frames here this is all

1:07:54

feature space of law this

1:07:59

is all feature space for Kent ex video him why would you know who you if I'm now look at the respective for myself ignoring the Red Cross this these are the Vaughan noise sells for the beauty of UK and if I'm not look at the broad noise sells for the X video ignoring the Blue depths these other Roma souls alert across the UK and that what used the cost that contains a shops from the the video well with 1 plastic You because this is more of a maximum long I'd with 1 custody and 1 custody of OPEC if there is a cost containing shops from to videos from the different videos the and by Wycombe this clustered by intersecting of lawn or a regions of the borough's some frames so while it after do at intersect this area over here and with this area over his head of Kent which he owes basically the area of the UK and a great shape area for these last this down you they don't contains any closer from the White City and they don't contains and a friend from the X video so nothing to be done about and what we see now is but the intersection of the space for the cost that contains framed from both videos is a good indication of Hull because the media celebrity between them

1:10:32

so the value office in this section is a measure of the video and it well there are some some some technical problems the for myself my measurable so we must have a way of measuring all features based on what we have to do 2 to consider all features space as compact so it's not unite this summer and on the other track friendly but it has a range between 1 and the role of and NHS and we need to know molestation so 0 justice to the total volume of defeat to face and the fraction of bullion that is occupied by just having frame from both videos is this kind of took the defection of the

1:11:23

videos and good sense are cost them law myself do not allow the video Siminovsky is basically so we will call it a lawn or video Samarra TVBS is basically the volume of the of the section which means as we have said that basically if we find 2 frames from different videos with the distance smaller than absolute we will just taken their individual bomb myself and want for Kent and take the union's of all these costs and get 1 of the few things they do not overlap for allways strictly separated the volume is just the some of the individual what you can sell the volume of some of the some of Williams and this is what we need to do we need to find the volume of the intersection of the different Clusters contained both ID and just summit at and the more cost us we find containing both the bigger other or not regions they will be and the because the volume of the Left will be and means that we can end up with a good videos

1:13:00

but in the summer we get a out-of-order videos of their T like for like we just did not like to see this as well in section of about a 3rd which is also consistent with the idea that the of the America's Cup the idea of video severity was again it 1 plus 2 we to dustily we clusters of lost 1 3rd of Kent yet from that does not only have to be the case because the good correlation between the ideal videos Murphy and and OR annoyed that system that he had stemmed from from the palm from from the fact that the all the deal 2 frames of Distributed quite evenly although the space so it really it makes them up a lot of the things I'm and and

1:13:58

and that may be different the oil was to be the case but what we want it could get for the board nuggety summer is not a random sampling and that is a different ball why we did the Lerner be American because we didn't want account Clusters because then we would have to count all the passed between frames and in both videos random Assembly work but if we use the full space than random samplings in the space and see whether they full into some shed reason between the video for some non shed reason between the video that gives a good impression of what I'm going to do is basically I'm going to have to feature space and going to a pub or club put and the videos In the Borat for my tessellations or something like that I'm not sell and a 2nd 1 for the other video like that of Kent and now I'm going to a two Pep of the routine with random sample points and 42 the random simple points I'd determined was suspected the videos is that in a shed regions or is not in the ship which and the number of points that I'm them using this kind of of cost correlated with a arrayed that I'm inquiry but just use 1 random sample point and its at so rather simplistic you not you is in and out of such a regional as not and then be summer to will be 1 0 0 0 the to make sense enough but it might take a thousand of these points of get a very good impression of Hall because faces a off because of the number awful for of she fastest with a of it this is a shame because this is a shame off and maybe this is a chip some of the of the of since it separated but it completely politicians the space the probability of getting it was a random see the remnants of a sample point from is directly propulsion 0 2 volumes the health was spectacle took space the more she shed space there is the more into sections them off the large of the space will be as opposed to the total Williams 4 1 look at this is the basic and the behind and and of shooting at this at the space with with the points and scene where they are located Icom by computer so this autumn to do and going to generate called factors that independently and uniformly Distributed all the space and the and a cheque for each of these points whether it's located in some intersection it there will much it is located in side some into effect you it has to be wrong sell of X and of why such that the 2 frames at the centre of both for myself are very close to each other more than the sum of and and whom them might take the frame from the X video that has the smallest distance to my seat point and the frame from the wide videos that has smallest distance to my seat point from end it is in the intersection musty points in the intersection the of the distance between these 2 points a small of a M Rees take a features PP take a seat point the 1 for you to see the point look at the closest point from the X video x 3 for example and broke the closest see point from but why video apparent them if the distance which you between these 2 with is smaller than Epsom as a 1 0 in and overlapping regions guest if the distance is the of that then these 2 different last as long as not in an over that things if I'd do that for a last point is 1 that are randomly and uniformly Distributed all the features space by would get an impression of how many intersection areas play off this gives me a sampling in tuition about the broader is emerging and thus to the ideal of the of UK and this is a base the idea that what really doing

1:20:11

now so we described each video P through the top book with respect to the embassy points Red tape the frames from the video P but I'm no simulators to my seat points and this is what we call the video some so that are powers to that EU the to the random something that we tried before along the video him up just take some random frame some excellent just take some random friend from white and compared them with in make much them we now although the video signature will say well take was range from X the closest 1 of the point and take those frame from wide the closest once he point and now we can competitive because we have matching has already was respectively see of this is the idea so and so but this is usually called a video signatures to every video gets a signature and and them in NEMO signature consisting of the frames closest to the sea but the UK and the celebrity that measure way up to videos is basically the degree of overlap between the respective signatures sell the phrase into signatures at the same position so with respect to the same teach the secret of his smaller of that by counted if it's not of the maximum different classes no over the UK and don't count divided by size of signature this is my basic foreign oil Simenon she and it has been retrieved by a random some I'd have abstracted every video into M frames but the obvious question of hell but them only we come to Latin America I'm but that is at least a possibilities of doing something that is at least a possibility not of looking at the pale wise simulate Ritchie's between between all the cost of all the things and

1:22:51

so that is good about this is the basic video signature Simoneti time because it is kind of like the based on this video statement and the basic because it has some of them that we will deal with later I'm since the seat Vectra's I'd uniformly distributed the probability of it is in the detection of 2 more noise Celso the not of across the road basically is the volume of losses directly with a is directly proportional to the with the volume that is made of by shed at cells as opposed to non shed OK I'm and that it is the Noibi of their teeth so the video signatures American is directly propulsion with the following the signature of 2 abroad or hideous and end of calls over the video collection you have to use identical see if to decide to randomly draw seats 1 time and then the abstract every video below the segment of every video in your that she was suspected exactly the seeds otherwise you come from of a video of the yes it doesn't matter but since you owe seeds are kind of Distributed even the all the features based on the unit for new distribute all the features of this rather rather improbable that he had collected but if they are in the same for myself all the videos and the seems to be a lot for myself and about large for not as good for the overlap which is good for them for the system to that found the

1:25:05

number of seats were just the kind of of on during the of calls we need a couple of seats just a single see this not enough but we need a couple from the home made it obvious that the lot of the M the more accurate the estimate that gets better if I'd take more samples have the morally M the more easy is the signature calculation defeat tomatoes although before the videos becomes more as told the BBC that the seed action of the correct M is kind of based on the on on on the air of ability are sold what should ICI electricity and Hall feistier the of pro ability to take a video database with and videos and and seats and I'd take some some some well constant that I've load defined maximum Khaled correlation the probability with respect to M is basically some the probability that the database contains the sick couple videos for which the difference between the law and order video Samaritaine and the media signatures American basic to signature of the memory is greatest and some of them some comes up and so so we don't want to much deviation between the 2 so we fix it was some constant gamma of and what we want to know as the probability that this

1:26:46

is respected by the videos and what happens is that the probability that the video of or not videos temerity and the beauty signatures to I'm deviate from each other more than this Gomaa with respect to the players of videos in the database and that is what we have to work with to compute and for a can show is that we if we choose M large of them logarithm of the number of videos and the collection of minus the of intended error provided by the standard deviation this will be sufficient and accurate to

1:27:42

prove that so if I'd take the wrong audio Simoneti and the Signature teeth that is dependent on the M obtained by can use half things inequality which basically give me the maximum Paul ability that some of independent random of the area but there the it with more than a given constant from expected valued by using the exponential which can be done from is a statistic a

1:28:15

measure then say while the deviation between the 2 full of the videos and the set is basically of the some of the individual of probabilities because the probe annuities are independently distribute it doesn't if 1 player is is a lot of and Michael not it has no influence on other peppers and assuming the independence of we know about the independent can just take the union and make it into some sort home than we can kind of Take That with costing inequality into the exponential functioned and hominy head of videos to we have in and all collection well we basically have to the of 2 5 2 so we have been reduced from a collection which makes and where it has and will have to take a pay off then a sufficient conditions if we take effect that so this is about the you are appalled that she and the other probability should be smaller than some of this is the maximum mistake we want to make a 4 to win the and if you want to be and we just have to sold the Beijing with respect to M which will basically you the error terms that showed before were for this is a good way to do so that a good choice of M is directly correlated to the number of videos in the collection and so to the error that we want to make and inversely correlated to the Mohmet discretion of the celebrities from each other for the close of the celebrities are the smaller my and and the last videos have the smaller might and in the large of the error by Idei might make the smaller might be if wants more LaRose races increase the and if moulded increase the and folk and is the basic anti behind

1:30:54

but the bomb SIAM and is a good thing as logarithmic in the size of the law group is 1 of these typical functions which are so it it it doesn't deviated the and so it goes to infinity but various slowly and that it really doesn't Mannerheimintie video that take it only growth various load of the good and terms of calls and the smaller the or the village are of all 4 of the mess that is the great to the values for which got UK and that so maybe we

1:31:44

exactly may we should make small break the 5 members and then

1:31:52

we go to see cryptogam regions of the UK

1:31:59

and so it's gone I'm at

1:32:06

pointed out before the videos of the law might be disability is not fully is the same as the idea of the American so we do well and that we were dealing with the with the case the what I'm it it has something to do with the evils of the state trains distributed well over the future space this is a case if the frames are kind of like by his in the foetus if this kind of not the case and he

1:32:36

is an example we can see that this is very of not the case so for example via the case with them the ideal of the American of 1 3rd of 1 parent of matching images and 2 clusters of non matching effects then that could be to connexions it could be the like like the and in the image and the space covered by the bomb ourselves is very small or it could be like you 1 trusted to justice because the again ideal devious and the of 1 3rd how the the low-noise video somebody quite because the videos to not use the whole features space but the rest of the foetus face is so to speak that is that empty but it is kind of up for grabs the for a cost of all foreign on clusters and that the fact that the run on the day of

1:33:41

the match so I'm what we wanted to follow for the rest of the lecture is we want to estimate the idea of video Summerteeth the basic video said the signatures American even if the overall Noisome Loti and the idea but his American before and obviously what we have to do it we have to hand over even or to dismiss match in space are so we have to we have to look at the density of the space and cut out at the non densely populated and so that they should not be re guided the valuable for for the estimation so I'm via the seeds are spread evenly for the future space but the estimation you of whether matches as the last the Quentin both millions of a classic containing just 1 media is obviously in front by the by the promise of a raw myself very big it will have a big wait for a big influence on the American list of I'm so where we have to do it we don't have to distribute the seeds even the although the space but rather have to distribute the seeds even the although the Vaughan myself so for example you

1:35:18

either take this example from period of what happened with my seats if I'd distribute them even for the space Eigo like this I'd take the but the pan over here and go throat and throat and look a case this evening is to be available space and randomly assigned how it's all the same bombers because the same have a lot of empty space where she should be doing is kind of a should Distributed the seeds even the although for noise self so they probability of hitting this long myself should be exactly the same as the probability of getting this for noise of all this for a bucket without considerate full because they actually or if it is to be even the although the space the size of the politicians would direct the influence the probability of getting a seat at last if I'd look at the politician in terms of lawn ourselves only the number of foreign myself should affect the simulated of the probability of getting a seat at the more or myself sell less a bility for each individual for myself a precise should not matter of care guests while is exactly the problem is that because we have to say we have to take the same seemed to letters for all the different video so we have to decide for a good week for last it's to we should take the via myself for each individual video and that will be the best way of doing but we kind do that because then we can of compared the deals any more because of different videos would result in different seat vector so different seat victims would result indifferent signatures and the match another thing but no longer be a sensible sell what we need to do it is that we will come to that in and the and we need to the site for 1 distribution of the director that is not to bed with respect to full of different but the with the videos in in my collection usually the this space free but and it's a good idea not to put seat factors into left space but you will talk about the later

1:38:15

so I'm the idea is really distribute the seats in the over or not cells regardless of

1:38:20

volume and the to generate steps each week but we confuse the unified distribution over after NEMO also this is not what not of that but we have to use a distribution with some density fact that as the idea that we have to videos and the Distribution density you found that at some point and the and and in the space is basically bomb many costs of their off and the value of the volume of sales has to be taken as a novelisations by the by by the volume of the law not so that means the probability of getting a seat in the cell is dependent inversely 1 home minicells them off but is well so you can or should be depend on novelization by the volume of this such that the biggest cells don't get more seat factors in the smallest of debate let less effective so I'm Nnamdi for every last the and the and the eyes of the law myself by William and an although as the total number of foreign ourselves for the basic probability of his is the idea behind while entry was the well

1:40:02

good like that it is inversely proportional to of the volume of the cell so we get a unifying distributed on the set of clubs every cost has the same chimed of getting a seat victory not counting the size of the for most of around and if we have of or myself after the density is a concept that was not the does not changed so we haven't equal distribution within the cost so really Keneally's is a weekend just of randomly which was Aclasta but which was a random point with the last but this random point was in the cost is not randomly Distributed all the space and the moped the move would be affected by the different Williams of different from ones but it is no life by the size of its the same that it was the same probe ability for every for every room that but if we

1:41:10

do not want the uniform reproduce seats but we want the density fact than we can't estimated the ideal video summer to you as the number off Fastow's those of a full number of over that from with respect for the unit for me all the cost of Distributed seat W this IS the the intersection of the cell were cell leaders of how many times with the key to victory by end up in a last the is still that it has and eye count the ones that are and acts on a pop at the numbers and UK and so for the new generated see points and look at of are myself at that over the and look which over that is generated by the common Class a between the videos and which over that is generated by a single video and and and found only the ones that I've been created by common friend of the media and that he is the cell if we take this as being the characteristic functioned so you for distribution all that this is exactly what said that was the for noise video similar to UK and that if we use to smooth density here that is long life by the cell sign that this moment the number of left them we got it OK I'm

1:43:12

the video signatures to simmered approximate the ideal videos the chief of the cost society the identical all very well but separate so what we do is that we have to be Dios such that to for the pay off classes index and and why either the classes are the same don't think after their over frames In The X custom off further away in lot of excellent so further the way it was suspected excellent from the frames in the test tube Clusters or you the and action on the part foreign they are the same and just 1 because of the way this 1 0 wanted and at Dawnay classes that that go like this the of very close to each other because the and what happened to secret review the could be classed as a country in both frame from both videos but it could well be that this is the closest here and this is the closest he this is not for the want of an eye was a OK their lollypop so if I'm somewhere between you Agate either represent of this Cluster all representatives of the but not of the basic than the ideal be just emerging is the density functions and the ball was that the size of the over that the volume of the over taken full what we already a

1:45:18

hat now we can show that the for each term in the time it the frames are just at the moment part than they must belong to the same trust this week assuming the glasses to be different it's not but it can canopy the case that a point in 1 Costa and a point in a different class of far smaller than that of UK and an another at the just exclude the case they must be to the same class sense and they belong to the same class psyche industry Fridays and the UK at this is basically the same pasta and this is exactly the part from of the of the meeting at take all across pop and the intersection between the clubs but this is all the Paris of frames having smaller distance and acted it

1:46:24

overnight sell them for full though for although frames that that we have been Acosta so we have the cost of its kind in the section of of of different Fastest just means that the into section is nothing but the value of the cell if I'd take both closest to invite them to a point and the for noise self the suspected BigSim Vaughan myself a suspect to to that the neck and take the Union of around sales of this was respect otherwise how would result that would end up in a different class that it would make sense because that is more than exit the time what and then do is kind of like a can on your since the of the same anyway this in this section here is basically the union of the cost but

1:47:29

and that means that if I'd put in the density functions no life by the size of the cell and no by number of to sell well this obviously has nothing to do with the bomb myself stoic brought out and this means that we have some of up the I'm all the Davies the and integral was a characteristic function divided for of Kent and was part of a year from this is all we have to be the size of the number of last that happened intersections with of drift from both because this column swanned where every last from both videos and this no mobilises the sites that doesn't monohull because cell as a discounted UK and leading so the number of intersections you by dint by the number of all the classes and now we have a set of similar Gusteau's and the head of all classes and which means that the set of Fastest containing from those with new divided by the number of cluster their which is exactly the idea of against APEC so if a you that this city French Nong lines and the generation of seat factos by a 3rd size of row myself I'd endowed with the idea of video and and this would just approved for care the you don't seem to be to combine the but the

1:50:01

so I'm and or comes comes from September Raymond basically because obviously it is not possible to use the same density function for the calculation of of producing signature because the density function different for every bit of it takes the value of all the different for myself into account the law noise cells Defoe for every video salsa low-volume stiffer give as a set of density function 1 for each video and that the makes sense because the for the comparison of videos the same seats must be used of West of frame stone match so I'm we deal with it like it's usually dynamic and that the area will just say Well lenses humour or collects is not who had urged that it's not totally different videos of based has some and there will be some general characteristics in terms of exploiting the features and these general to characteristics can be learned by just taking over representative training sent and determining the density function for this training set and then the using the single density functioned full of videos then we don't have a problem any more than we have different density function for different videos and we can competitive and the mistake that we make sense for used a representative said that somehow kept the general characteristics of all of the deal collection is not to be again we are not we are not in the ideal videos and case but we are close to the idea of the mood of this

1:52:11

is for you can do so and the UK with the for generating single and we just do this and times to get and seats is we use some value Epsilon that is of the state sector and the a generation and the training set of frames which restrict the collection as as as well as possible for so this kind of the idea and we identify classes of the set and then to use any of the clubs and and discussed we will put the by train in the UK and so the idea is basically which was the value that is all well excellent value but choose a training sent that was used to build all density functions and then we identify Clusters was suspect to this absolute of the trend except the random which was any of the fastest small maligned the anti Vaughan are still be generated steep point for that but

1:53:40

low after we generated the sea point raw myself we or and generated random Bektas all the foetus says well until 1 of them is in the ball myself so I'm the idea of how to do that I'd do if I just based on idea have fastest in the foetus space might have different the size for myself but public very small this 1 for the big but you to my clever no Molise a due to my clever moment as a suit the probability of choosing discussed is exactly the same as the pro ability of choosing discussed of his because Naum not off dive for example said on this Costa the cost of wanted by John Reid the centimetres Rossi to victory in what led to well eye can generate random because all the features based just going back to pump and pomp and palm and as you has some life into the cell that take that 1 of the disease this way of for of generating found well 1 could also use a random frame from this classed as a seat at just look at the cost and take any random random random image of the club said that this is the see so the seat to value doesn't have to apply life somewhere in the world order regions but it can also be part of the cost to possibilities of actually creating the seat as the the randomly where a look at the read the myself or just take note of the cost but if

1:55:48

we do that and and experiments on so for example we have 15 million from the impact of the continent that would just kind of like a typical test said eye and and we create a new videos by a random the snow frames to such that we have different video game that different ideal videos tomorrow to use the old 1 8 0 6 0 0 4 4 4 2 experiment tried to to determined the video signature symplectic and generated a 100 seats for that you need if only distributed on undefeated based all based on the text collection of of Fort of photographs of the and reflected the Ampex content and what we can see is that are full of day the idea of video Simoneti that was used of all point 8 0 points 6 0 0 4 4 0 4 to compel with the simulator T are by the unit for new Distributed seats and Distributed seats with this but density function used from with the training for use was definitely much bigger aviation's for the beautiful only distributed features and much smaller than the deviations for the train to density for which that it seems to be the case at least for this for this collection in which is kind of a very representative in Brixton companies that could videos of all kinds basically and it seems to be the case of that really some part of the features that I'm not really exploited by most of the ideas and the and normalising the density function by the size of raw myself to help them even if it's just the way

1:58:08

represented the foetus at that I'm so for what we've done and UN to the now is basically which look at the basic beauty of signatures of her to and we looked at the idea of video celebrity and we've on Cuba well basically it's the same because the video signature simpler to reflect some of whom the simulated if we normalise by the number of homes although the size of the overall noise but we can do but we did make a 2nd assumption that we made this time she that the cost us by well separate that these maximum a part side don't get to these these these items were found on of of and 4 0 in the 2nd time this gap and say all this is the closest from here and this is the closest from the they belong to a different of the belong to different class this whole ever they are only Apslawn a part of Lloyds on during that could well distinguish last us by looking at the distance between 2 trains and if they were kind of Madeleine part they had to be launched before and if they were small than that of they had to belong to the same trust and this is not really true from so if the cost of identical to be separated this is definitely but the feature of calls off in and the and the approximation off the visual perception that we are and there are some MPs some discrepancies so it it can be that usually similar last or close to be of and this is

2:00:11

often seen as you know if we have a video for example that is slightly tilt at of a change of colour to in the site there relations so I'm aided by making the right of life at more contrast was still the same video it has still to the same visual impression of the future of will change slightly what happens is that the point and space was shifts like and though the on 3 ideal videos to match definitely is 1 because they will cost us that have contained video Cook on frame from both videos the Vannoy diagrams off slightly different because this Costa has a certain Diamanta that shifts of all my diagram and death or the CDU because gaps over here exist because this is the cell for the wreck point and this is the cell for the group cross there are about the same but only just so slightly what happens now is my seat where should follow into 1 of these places and what happens if what other closest frames of each media I'd assignment but it full recommend your the row noise sell Estonia so the video signed saltiest the the 1 over him for the Blue video the Roma sell up there are still those who won the signs of his 3 of their the seed of a sign to different class of though there is a cluster of that contains both friends and just because of the frames of not identical the best slightly tilt at her off process bat because as soon as we have a seat being pollution and and 1 of these in 1 of these gaps began wrong all of them all measurement

2:03:27

of the space sold since the videos of the vulnerability of simulated the them is defined by the similar not region by the set in the fictional for my recent that following the temerity is strictly smaller than the idea of video some it does not account for the gaps and I'm you can Calculate this this difference by but it by the offset of 3 space but the moral detailed is bigger the cost of off on the more free space he will happen that means the bigger mistake he will make it a

2:04:10

so if we consider seed between the bomb cells long but we call this of for my gap from this is of for my gap I'm band exactly what happens to it happens were what I'd just said in all like so this is a sign to this and this is a sign to the US and both of them are large and that they do not belong to the same clubs and this is not affect because they both have correct substance intersecting aren't so I'm what we should do is we should of their way

2:04:52

to gaps of which should OID gets in in considering the seats we should always put the seats into 1 myself not between were it to avoid the death for for seed from the generation to fill the void or my gap for to videos is basically the said of all the features like those where the difference between the the next frames were suspect to X video and with which to buy a video large of the maximum but there is a frame in excess that is very close to the assign Freeman why and there is frame and white and that is very close to the bomb assign frame acts of Kent was exactly what would be before sony is also seek victory and we have the expertise to here and the wire to here so the is G white campuses she X but we have some 1 the here and we have some expected here and they blew it just a correct but was the case of the from book and that and

2:06:27

we want 1 can show and we need to experiment that of full for simple features basis in complex based at if it deeds out some but if I of sympathy to places like allies to prompt the most effective of some of very of use of the and the Vannoy gaps although the VNR incurred by the or my gap is considerable are so it can offer the neglected and usually have seats at the whole into the wrongly and fastest told the estimation of causal smaller the X along the but close over Clusters off the smaller before annoyed gaps sold what we use 0 1 to do not all as the last of lecture is we want to we want to avoid is the use of seats that well at least for some ability Lyon of long ago are so if

2:07:23

we were randomly generated and seats of which and life in the of the video signature simulator T of the Romanian and minus and victims is exactly the idea of the because their high don't have the problem of the signing from last so I'm but it can we some half find out whether some seat is in the law my gap all not and the answer

2:08:01

is part of that the definition of lawnmowing capped the not with the help of rape occasion because you you basically after 2 0 2 2 2 a distance calculation between each signature wept as an or phase of the of the video Moon so as to win As they want to have a seat at some point and find some some some value from the expected but NAO would have to do this and calculation to or what 1 9 and for to wise 3 year and so on but to find out whether there a single why am where this small of of the so called 42 signed point away to a televised look up of all the frames of the of the video to determine whether there is something but that is building the trust was not a good idea buy and the and it is right to set the video I'd just 1 to 2 have these these frames acts are full the video signature with respect to the M seats when and needs a lot of time to calculate the signature the use of or basically in the early days the the something Iike to found that the M of cost 1 could say but there are certain probabilities for the fact that the seat is a forum akin to say what I'd on to read Chadian out adjusted to find a probability distributable of space that tells me whether classes are not and where the probable regions of foreign oil gets

2:09:52

up so I'm if those video second half of the roughly at the distance paraphrased was suspected the seat for cell there is something in the video acts that is close to assign point and there is the frame and the baby white that is close to the White point it is clear that the frames as such by dissimilar from soul the seeds and the more Moika under the of different cells direct so and the Blues and and the Blues so here the Dome's the Blue X point and the rest so here be longs to red fixed point or so if I'd have an ex point and the assigned point of the and this is large of apsidal so they don't belong to the same class of their those from from video X over here and and over you but they don't belong to the same cost but the distance of the city point is somehow very similar police points them the see point period 12 ably is in along with the I'd take 2 points from the same video and they should be a part more of an excellent so they are not part of the simplest but they shared the same distance to some see what that means but it means that the city point is very pro globally close to along the way at because between those 2 point and X and similarly those who point and why de must be of honourable in the middle yes if I'd put the seat point somewhere very close to the long on board the chance so there is a gap by slightly the fault of the White was X video is very high if iconic finds the different point from the other video that is 1 of the criticism them I'm right in sight some self a print so the 2 cases like that we have and Expo year XFO you see point you can't and this is rather similar distances but these 2 would be long to different Clusters the back custody and direct last the and there must be a bomb noise distinction of 1 or of between the this the Blue will stretches here are very similar the 1 the chimes that my seat life is owned is high and if them the same holds for for some white values they will also have the ball and that the chances of life environment of the following camps are all lies in the vicinity of for not board keep Austique victims well away from those and we don't have the ball from this is the basic ideas but so

2:14:17

given to videos was excellent compact classes those of the for every seat in the following year get there is a victory in X and a victory in why such but is the act is this similar to the assigned as X frame for us and best the distance between the 2 ex frames large of a maximum but the distance between the ex frame and the point in the 2nd extreme embassy point this is very similar so we will says the ECB the distance if the if difference in this and his smaller than 2 at UC and good self can just prove

2:15:19

it since all point is in the world right that we have the difference between the point assigned to add the did so with the wrong again here the seat point here so the 1 Assign for the axe is bomb Gete X air for Kent and the 1 who signed for the wife if she was back of the and since the my gap the distance between the X and the White point as lot of that again but this is lot of but since the classes are by assumption that long compact the G excessive compute the same cost as the X so there must be some tax from over here and there must be some why point over the firm excellent compactness of the cost those a must be something wrong but and this is definitely book smaller at and this is also differently smaller than absolute that means the distance between those point is also a lot of land at the FSA but it though distance between the X and the seat point and the of time X embassy point for the distance between 2 this one off and the distance between this 1 or Kent this is nothing but a different yet OPEC this means by the trying to a quality a could also estimated this defence by going to the point why and then continuing to go to point X update 3 a new equity inequality must be large of this summer 1st going to point why banned from point white point and since this over here and now stands a we know this 1 must be of an absolute from a can just put in the extra low of the mountains and the sea and the

2:18:19

in the for my gap we know that there is a wide and the other video such at the White and be assigned point next smaller than excellent and we know that the same thing for the for the other possibilities these of the excellent context that some of these are the excellent complained that Clusters and this means that we can basically take the chosen point in Epsilons and the X along from and that's the point over here is similar to a point where you it must be further away so this is a what this small I'm we can estimate but she later this was assigned point of whom she had 2 of us and this is the other point that to land and if this would be smaller than she wide of less and would have been the assign point after has to be large of a pay so that the distance between white and that is all these large L and the distance between G white S and as a pin and that means that this is between the G white as the US and the distance between G excess and S is smaller than Exelon of because he you is also an ex value and again the tree Anglo inequality use that if I'd go there and then go there this has to be smaller to launch and this is G X best of Kent at that point in the summer my ex got lost best using 2 at so the difference between wide and G excessive it's more than that to write good

2:21:15

several OMX we can do the criterion from that some over and over again wrong but we just test whether the seed is in the following camp between some video and another random 2nd if there is no better in the ex video such that the X this amount to a frame chosen for this of chosen for the full seek victory that the distance between you know 1 the smaller than to add to then we can say the seat was not in the world like sort for every point X we assigned to all seat as she and we do not find any of next such that the distance for the difference in distance as a smaller than to excellent than we know this is not a neck at canopy it far away from the board of the visible we do

2:22:35

so we define a ranking function for the signature Victoria enter a you a possible seats by the members of of the distance yes we just look at when we have the minimum distances between frames of Act which to not belong in the same and the further away the seeds are from the board of for myself the high debt is the value of the Frankie function for care because the minimum of 50 feets high higher values of Hugh

2:23:27

are mocked and bright low-value of Hughes all my in black if we use the ranking from she and we can see the of whom ISA are basically made up by the Sri lines of so we can see this ranking function really computer but well with the error of its ways but something is in the gap all not and that is good because the gap that exists is kind of like given by the different wife that still belonged to the same Cluster but slightly off and and we want to do not take seeds from from this area is the black regions of all rankings function so by feuding all seats that have low-ranking function will immediately yet see points that this following from the possible for my gaps Boca the basic idea

2:24:39

behind so safety said he will use large and 2 excellent this is what we just proved and of course is not require also about within these 2 excellent which does not have to be an overnight but it could be so why should we respect of basic idea and it's not required but a sufficient and and and the and the and sometimes are a lot of general of many seats with fuel with solo laws to non not and long like that but we just don't care which is generated Mariah seeds and shoes only the 1 with the best he well that for Blue do so we definitely of what the from my gaps and that kind of time of track

2:25:25

so we do some seats than the actual size of the of the Signature should be we generated a set of and prime seat factors than compute ranking function for the seat factos Thirlwell the seeds according to the freezing Cuba use and take only the M 1st thus we avoid the borough my gaps never seems and from and those who need to look to the basics of the deal signature to we can not to find the ranked video similarity but we take the ones that are highly rated Hugh into account more than the

2:26:16

other way so basically the symmetrical on video signature simpler to with ranking function between 2 videos is defined by the sea was the highest ranking in White and S Byrne so I'd take half of them but Ohio's was suspected to the X video and I'd take half of them the highest with respect to the wife of the from but due to for both videos Toulouse the M of which part from bands well that just look at the different rankings of the Signature frame and although the rankings with respect for the few you value of only take the highest Hugo

2:27:12

use and found out that means that the system that you use 50 per cent of the firm's was the highest ranking was respect to the X video and 50 per cent some of the highest ranking the suspect to the White City and again by just end up with a M dimensional signature Cuemba that kind of some of close 1 can also use assymetric video video signature seminar to just using the M highest ranking the suspect 1 we deal told along both videos and and use them as a matter of than it would be assymetric

2:28:03

but by the estimated from lead to some distortion of calls from the of the the as the made if the video as a possible segment of another day of the assymetric videos Cygnet to some of it is still a significant behind when quite a calculated with the shoulder video and was along way because I'd have less degrees of freedom for Fault review the rankings will deteriorate more quickly and I've basically game a still have some before we get but the idea that if

2:28:49

you look at the retrieval and if activities taking the basic signature and the of ranked signature I'm than we see that decision re called for low ranked signature is actually higher than for the basic signature so we get the improvement from and firm before user menu evocation of and flow of well as wrong to troops from for the 4th is Rico and other this this also called for different sizes of of the signature of the on rank media Simoneti is definitely or with the and the basic and this this this is what we knew was basically by the wrong Mike at chemicals of a book that was before but

2:29:53

this lectured work on a like considering video simulator to sell good idea for you to bowl something like that to a fine similar videos and maybe just 0 1 lost all the ones with the highest quality of the ones that are best for the use of the appreciate the old of and also you and that you the pro just doing it just to computing the classes and looking for many shed classed do we and but that doesn't scale the and we can of randomise account on so that the 94 dissimilitude by of the big brand of samples so why we did is we get to the bottom ideas simulator to where we just as well as basic tessellation of of the different and friend of Frank classes and the intersection of the tessellations Laclos's containing frames of those videos interesting point this is what you want and then were account of defining under the I'm simonetta along the lines of if we randomly distributed the seeds of the wild of complete features days we might end up with many features a with many things in a single cell that is just very big that this was something we don't want but book what we rather would want is that we are I'm take the density function as a function of the law my clusters of raw myself not as the size of the rights as a William was so it should be the same Paul ability for every for myself non considering its its size to be hit by a feature of and the and the worst of clearing up the bomb ICAP's that are kind of like a shift in and the cost us and and that might have come through the fact that this is what we did today question is not members will inform developed simpler to get then next Sefton

2:32:11

that we will talk about the traction and we will talk about how to prevent videos of the and the of the retrieval a result of the deal such to the use that because that is obviously not as simple things would be due to use the new skimming highlighting with discuss a lot of possibilities to show the results to leave thankfully

00:00

Multimedia

Distributionstheorie

Bit

Programmiergerät

Subtraktion

Rahmenproblem

Multimedia

Euler-Winkel

Element <Mathematik>

Abstraktionsebene

Computeranimation

Datenhaltung

Benutzerbeteiligung

Informationsmodellierung

Einheit <Mathematik>

Statistische Analyse

Information Retrieval

Datenstruktur

Diskrete Wahrscheinlichkeitsverteilung

Statistik

Kraft

Exploit

Datenstruktur

Menge

Last

Hill-Differentialgleichung

Kantenfärbung

URL

Modelltheorie

Ordnung <Mathematik>

Schlüsselverwaltung

02:25

Quelle <Physik>

Rahmenproblem

Zellularer Automat

Zahlenbereich

Kartesische Koordinaten

Raum-Zeit

Ähnlichkeitsgeometrie

Computeranimation

Deskriptive Statistik

Informationsmodellierung

Einheit <Mathematik>

Gruppe <Mathematik>

Total <Mathematik>

Information Retrieval

Inhalt <Mathematik>

Einflussgröße

Nichtlinearer Operator

Shape <Informatik>

Codierungstheorie

Softwarepiraterie

Bildauflösung

Ähnlichkeitsgeometrie

Symboltabelle

Visuelles System

Frequenz

Quick-Sort

Rahmenproblem

Flächeninhalt

Last

Analogieschluss

Nichtunterscheidbarkeit

Ablaufverfolgung

Aggregatzustand

06:17

Resultante

Information Retrieval

Systemidentifikation

Distributionstheorie

Rahmenproblem

Mathematisierung

Zahlenbereich

Term

Ähnlichkeitsgeometrie

Computeranimation

Textur-Mapping

Multiplikation

Zeitreihenanalyse

Information Retrieval

Distributionstheorie

Shape <Informatik>

Teilbarkeit

Prozess <Informatik>

Systemidentifikation

Systemaufruf

Einfache Genauigkeit

Symboltabelle

Rechnen

Visuelles System

Reihe

Rahmenproblem

Fundamentalsatz der Algebra

Minimalgrad

Last

Mereologie

09:14

Relationentheorie

Subtraktion

Bit

Punkt

Rahmenproblem

YouTube

Familie <Mathematik>

Gruppenkeim

Zahlenbereich

Kombinatorische Gruppentheorie

BAYES

Dicke

Ähnlichkeitsgeometrie

Computeranimation

Homepage

Benutzerbeteiligung

Subtraktion

Warteschlange

Operations Research

Sichtenkonzept

Softwarewerkzeug

Elektronische Publikation

Dateiformat

Verknüpfungsglied

Benutzerschnittstellenverwaltungssystem

Basisvektor

TVD-Verfahren

Parametersystem

Ordnung <Mathematik>

Einfügungsdämpfung

12:36

Größenordnung

Relationentheorie

Resultante

Bit

Rahmenproblem

Klasse <Mathematik>

Selbstrepräsentation

Gruppenkeim

Vektorraum

Green-Funktion

Zahlenbereich

Computer

Framework <Informatik>

Code

Raum-Zeit

Ähnlichkeitsgeometrie

Computeranimation

Freeware

Benutzerbeteiligung

Nichtunterscheidbarkeit

Abstand

Cluster <Rechnernetz>

Fehlermeldung

Extremwert

Vektorgraphik

Ähnlichkeitsgeometrie

Elektronische Unterschrift

Arithmetisches Mittel

Diskrete-Elemente-Methode

Hypermedia

Selbstrepräsentation

Ruhmasse

Quadratzahl

Information

Euklidische Ebene

Fehlermeldung

16:24

Rahmenproblem

Vektorraum

Term

Whiteboard

Raum-Zeit

Ähnlichkeitsgeometrie

Computeranimation

Unendlichkeit

Softwaretest

Algorithmus

Abstand

Einflussgröße

Folge <Mathematik>

Videospiel

Vektorgraphik

Matching <Graphentheorie>

Raum-Zeit

Zwei

Linienelement

Symboltabelle

Wiederkehrender Zustand

Elektronische Unterschrift

Menge

Invariante

Rahmenproblem

Selbstrepräsentation

Ordnung <Mathematik>

Standardabweichung

19:48

Information Retrieval

Lineares Funktional

Dicke

Total <Mathematik>

Gewicht <Mathematik>

Matching <Graphentheorie>

Rahmenproblem

p-V-Diagramm

Zahlenbereich

Ähnlichkeitsgeometrie

Ähnlichkeitsgeometrie

Hinterlegungsverfahren <Kryptologie>

Computeranimation

Rahmenproblem

Perfekte Gruppe

Funktion <Mathematik>

Prozess <Informatik>

Total <Mathematik>

Hypermedia

Abstand

Inhalt <Mathematik>

Verzeichnisdienst

Hilfesystem

23:29

Abstimmung <Frequenz>

Wellenpaket

Multiplikation

Rahmenproblem

Symboltabelle

Physikalisches System

Quick-Sort

Raum-Zeit

Ähnlichkeitsgeometrie

Computeranimation

Rahmenproblem

Abstand

Fitnessfunktion

26:35

Subtraktion

Merkmalsraum

Rahmenproblem

Momentenproblem

Klasse <Mathematik>

Gruppenoperation

Mathematisierung

Gesetz <Physik>

Ähnlichkeitsgeometrie

Computeranimation

Graph

Komponente <Software>

Kompakter Raum

Inverser Limes

Abstand

Cluster <Rechnernetz>

Datenstruktur

Widerspruchsfreiheit

Autorisierung

Oval

Vektorgraphik

Güte der Anpassung

Ähnlichkeitsgeometrie

Grundeinheit <Mathematik>

Ereignishorizont

Auswahlverfahren

Rahmenproblem

System F

Minimalgrad

Verkettung <Informatik>

Datenstruktur

Menge

Einheit <Mathematik>

Kompakter Raum

Selbstrepräsentation

Ultraviolett-Photoelektronenspektroskopie

Klumpenstichprobe

32:52

Programmiergerät

Subtraktion

Einfügungsdämpfung

Punkt

Extrempunkt

Klasse <Mathematik>

Mathematisierung

Snake <Bildverarbeitung>

Element <Mathematik>

Term

Ähnlichkeitsgeometrie

Computeranimation

Komponente <Software>

Spannweite <Stochastik>

Gruppe <Mathematik>

Total <Mathematik>

Datentyp

Kompakter Raum

Optimierung

Cluster <Rechnernetz>

Ideal <Mathematik>

Videospiel

Matching <Graphentheorie>

Quick-Sort

Ereignishorizont

System F

Rahmenproblem

Mereologie

Ordnung <Mathematik>

Klumpenstichprobe

Verkehrsinformation

39:34

Resultante

Subtraktion

Bit

Total <Mathematik>

Punkt

Rahmenproblem

Freeware

Klasse <Mathematik>

Zahlenbereich

Gesetz <Physik>

Term

Ähnlichkeitsgeometrie

Whiteboard

Computeranimation

Spannweite <Stochastik>

Arithmetischer Ausdruck

Total <Mathematik>

Cluster <Rechnernetz>

Gerade

Metropolitan area network

Soundverarbeitung

Lineares Funktional

Güte der Anpassung

Ideal <Mathematik>

Rahmenproblem

Ordnung <Mathematik>

Klumpenstichprobe

45:26

Mittelwert

Resultante

Total <Mathematik>

Merkmalsraum

Rahmenproblem

Klasse <Mathematik>

Selbstrepräsentation

Voronoi-Diagramm

Diagramm

IRIS-T

Zahlenbereich

E-Mail

Rechenbuch

Ähnlichkeitsgeometrie

Raum-Zeit

Computeranimation

Gruppe <Mathematik>

Schätzung

Stichprobenumfang

Abstand

Cluster <Rechnernetz>

Figurierte Zahl

Gerade

Schreib-Lese-Kopf

Dicke

Matching <Graphentheorie>

Stichprobennahme

Stichprobe

Symboltabelle

Ideal <Mathematik>

Ähnlichkeitsgeometrie

Routing

Rechnen

Arithmetisches Mittel

Rahmenproblem

Mereologie

Ablöseblase

50:22

Ebene

Mereologie

Punkt

Voronoi-Diagramm

Diagramm

Zahlenbereich

Raum-Zeit

Computeranimation

Demoszene <Programmierung>

Spannweite <Stochastik>

Parkettierung

Prozess <Informatik>

Hyperebene

Datentyp

Punkt

Abstand

Taupunkt

Mathematiker

Mathematik

Matching <Graphentheorie>

Green-Funktion

Raum-Zeit

Division

Linienelement

Hyperebene

Systemaufruf

Physikalisches System

Dialekt

Quick-Sort

Bildschirmmaske

Diagramm

Datenfeld

Rechter Winkel

Mathematikerin

Heegaard-Zerlegung

Helmholtz-Zerlegung

Ordnung <Mathematik>

Aggregatzustand

53:45

Resultante

Bit

Punkt

Voronoi-Diagramm

Diagramm

Kartesische Koordinaten

Computer

Computerunterstütztes Verfahren

Login

Komplex <Algebra>

Analysis

Raum-Zeit

Rechenbuch

Computeranimation

Hyperebene

Abstand

Maschinelles Sehen

Algorithmus

Sichtenkonzept

Hyperebene

Paarvergleich

Dialekt

Ebene

Automatische Indexierung

Konditionszahl

Zellularer Automat

Mereologie

Euklidische Ebene

57:11

Quelle <Physik>

Subtraktion

Punkt

Rahmenproblem

Klasse <Mathematik>

Voronoi-Diagramm

Diagramm

Geräusch

Zahlenbereich

Abgeschlossene Menge

Computer

Gesetz <Physik>

Ähnlichkeitsgeometrie

Whiteboard

Raum-Zeit

Division

Computeranimation

Schwebung

Gruppe <Mathematik>

Mini-Disc

Abstand

Merkmalsraum

Cluster <Rechnernetz>

Funktion <Mathematik>

Lineares Funktional

Vektorgraphik

Raum-Zeit

Division

Verzweigendes Programm

Ausnahmebehandlung

Mailing-Liste

Gleichheitszeichen

Dialekt

Rahmenproblem

System F

Diagramm

Parkettierung

Zellularer Automat

Mereologie

Bus <Informatik>

1:04:33

Subtraktion

Gewichtete Summe

Rahmenproblem

Flächentheorie

Voronoi-Diagramm

Diagramm

Abgeschlossene Menge

Zellularer Automat

Geräusch

Gesetz <Physik>

Punktspektrum

Ähnlichkeitsgeometrie

Whiteboard

Computeranimation

System F

Flächeninhalt

Zellularer Automat

Mereologie

Nichtunterscheidbarkeit

Hypermedia

Flächeninhalt

Nichtunterscheidbarkeit

Abstand

Spezifisches Volumen

Klumpenstichprobe

1:07:39

Relationentheorie

Bit

Merkmalsraum

Rahmenproblem

Voronoi-Diagramm

Geräusch

Gesetz <Physik>

Raum-Zeit

Ähnlichkeitsgeometrie

Computeranimation

Spezifisches Volumen

Unendlichkeit

Volumen

Flächeninhalt

Indexberechnung

Merkmalsraum

Schreib-Lese-Kopf

Shape <Informatik>

Güte der Anpassung

Dialekt

Rahmenproblem

Integral

Flächeninhalt

Zellularer Automat

Hypermedia

1:10:31

Relationentheorie

Subtraktion

Total <Mathematik>

Merkmalsraum

Rahmenproblem

Voronoi-Diagramm

Gesetz <Physik>

Ähnlichkeitsgeometrie

Computeranimation

Spezifisches Volumen

Unendlichkeit

Spannweite <Stochastik>

Volumen

Spezifisches Volumen

Abstand

Cluster <Rechnernetz>

Merkmalsraum

Einflussgröße

Bruchrechnung

Güte der Anpassung

Dialekt

Office-Paket

Integral

Kompakter Raum

Zellularer Automat

Garbentheorie

Klumpenstichprobe

1:12:59

Korrelationsfunktion

Punkt

Merkmalsraum

Rahmenproblem

Voronoi-Diagramm

Zahlenbereich

Computer

Ähnlichkeitsgeometrie

Whiteboard

Raum-Zeit

Computeranimation

Spezifisches Volumen

Demoszene <Programmierung>

Schätzung

Stichprobenumfang

Randomisierung

Spezifisches Volumen

Abstand

Cluster <Rechnernetz>

Korrelationsfunktion

Prinzip der gleichmäßigen Beschränktheit

Soundverarbeitung

Schießverfahren

Vektorgraphik

Stichprobennahme

Raum-Zeit

Güte der Anpassung

Pauli-Prinzip

Ideal <Mathematik>

Physikalisches System

Widerspruchsfreiheit

Teilbarkeit

Dialekt

Flächeninhalt

Parkettierung

Zellularer Automat

Ablöseblase

Garbentheorie

1:20:10

Schätzwert

Subtraktion

Einfügungsdämpfung

Punkt

Ortsoperator

Rahmenproblem

Extrempunkt

Klasse <Mathematik>

Zellularer Automat

Geräusch

Zählen

Ähnlichkeitsgeometrie

Rechenbuch

Computeranimation

Spezifisches Volumen

Spannweite <Stochastik>

Elektronische Unterschrift

Diskrete Simulation

Schätzung

Spezifisches Volumen

Ereignishorizont

Einflussgröße

Leistung <Physik>

Befehl <Informatik>

Vektorgraphik

Magnetbandlaufwerk

Systemaufruf

Physikalisches System

Matching

Elektronische Unterschrift

Nichtunterscheidbarkeit

Lateinisches Quadrat

1:25:01

Subtraktion

Extrempunkt

Gruppenoperation

Zahlenbereich

Extrempunkt

Gesetz <Physik>

Rechenbuch

Computeranimation

Datenhaltung

Logarithmus

Elektronische Unterschrift

Standardabweichung

Konstante

Stichprobenumfang

Schätzung

Rechenschieber

Korrelationsfunktion

Schätzwert

Fehlermeldung

Datenhaltung

Systemaufruf

Rechnen

Elektronische Unterschrift

Auswahlaxiom

Last

Festspeicher

Hypermedia

Ordnung <Mathematik>

Beweistheorie

Standardabweichung

Fehlermeldung

1:27:42

Extrempunkt

Zahlenbereich

Abgeschlossene Menge

Diskrete Gruppe

Extrempunkt

Term

Computeranimation

Ungleichung

Konstante

Schätzung

Konditionszahl

Auswahlaxiom

Einflussgröße

Schreib-Lese-Kopf

Soundverarbeitung

Statistik

Exponent

Stochastische Abhängigkeit

Güte der Anpassung

Elektronische Unterschrift

Variable

Quick-Sort

Flächeninhalt

Menge

Ungleichung

Konditionszahl

Fehlermeldung

Standardabweichung

1:30:52

Wellenpaket

Merkmalsraum

Rahmenproblem

Voronoi-Diagramm

Gruppenkeim

Vektorraum

Term

Gesetz <Physik>

Ähnlichkeitsgeometrie

Raum-Zeit

Computeranimation

Datenhaltung

Logarithmus

Schätzung

Vererbungshierarchie

Kontrollstruktur

Cluster <Rechnernetz>

Merkmalsraum

Soundverarbeitung

Lineares Funktional

Fehlermeldung

Systemaufruf

Ideal <Mathematik>

Dialekt

Einheit <Mathematik>

Last

Ganze Funktion

Klumpenstichprobe

Aggregatzustand

1:33:41

Distributionstheorie

Subtraktion

Web Site

Voronoi-Diagramm

Geräusch

Vektorraum

Term

Raum-Zeit

Ähnlichkeitsgeometrie

Computeranimation

Distributionenraum

Schätzung

Volumen

Merkmalsraum

Schätzwert

Matching <Graphentheorie>

Güte der Anpassung

Mailing-Liste

Vektorraum

Elektronische Unterschrift

Frequenz

Teilbarkeit

Dichte <Physik>

Zellularer Automat

Hypermedia

1:38:13

Distributionstheorie

Subtraktion

Punkt

Dichte <Physik>

Voronoi-Diagramm

Zellularer Automat

Vektorraum

Gesetz <Physik>

Raum-Zeit

Computeranimation

Spezifisches Volumen

Eins

Distributionenraum

Schätzung

Konstante

Volumen

Spezifisches Volumen

Gleichmäßige Konvergenz

Merkmalsraum

Videospiel

Distributionstheorie

Dämon <Informatik>

Gleichheitszeichen

Teilbarkeit

Dichte <Physik>

Menge

Zellularer Automat

Klumpenstichprobe

Gleichverteilung

1:41:05

Schätzwert

Distributionstheorie

Theorem

Momentenproblem

Rahmenproblem

Dichte <Physik>

Klasse <Mathematik>

Gruppenoperation

Selbstrepräsentation

Vektorraum

Geräusch

Zellularer Automat

Zahlenbereich

Computeranimation

Eins

Einheit <Mathematik>

Dichtefunktional

Vorzeichen <Mathematik>

Spezifisches Volumen

Gleichmäßige Konvergenz

Cluster <Rechnernetz>

Softwaretest

Distributionstheorie

Videospiel

Ideal <Mathematik>

Elektronische Unterschrift

Dichte <Physik>

Rahmenproblem

Automatische Indexierung

Mereologie

Klumpenstichprobe

Schlüsselverwaltung

1:45:17

Subtraktion

Punkt

Rahmenproblem

Momentenproblem

Klasse <Mathematik>

Geräusch

Zellularer Automat

Extrempunkt

Term

Computeranimation

System F

Zellularer Automat

Mereologie

Garbentheorie

Abstand

Term

Beweistheorie

1:47:26

Bit

Subtraktion

Web Site

Wellenpaket

Rahmenproblem

Dichte <Physik>

Klasse <Mathematik>

Selbstrepräsentation

Geräusch

Zellularer Automat

Zahlenbereich

Gesetz <Physik>

Term

Rechenbuch

Computeranimation

Datensatz

Dichtefunktional

Schätzung

Drei

Gerade

Schreib-Lese-Kopf

Lineares Funktional

Videospiel

Matching <Graphentheorie>

Paarvergleich

Ideal <Mathematik>

Ähnlichkeitsgeometrie

Paarvergleich

Elektronische Unterschrift

Dichte <Physik>

Generator <Informatik>

Funktion <Mathematik>

Menge

Flächeninhalt

Einheit <Mathematik>

ATM

Mereologie

Charakteristisches Polynom

Klumpenstichprobe

Term

1:52:10

Algorithmus

Videospiel

Suite <Programmpaket>

Wellenpaket

Punkt

Vektorgraphik

Momentenproblem

Rahmenproblem

Raum-Zeit

Klasse <Mathematik>

Zellularer Automat

Raum-Zeit

Dialekt

Computeranimation

Rahmenproblem

Generator <Informatik>

Zufallszahlen

Dichtefunktional

Menge

Twitter <Softwareplattform>

Zellularer Automat

Mereologie

Ordnung <Mathematik>

Cluster <Rechnernetz>

Klumpenstichprobe

Aggregatzustand

Grenzwertberechnung

1:55:45

Mittelwert

Subtraktion

Wellenpaket

Punkt

Rahmenproblem

Extrempunkt

Klasse <Mathematik>

Voronoi-Diagramm

Content <Internet>

Geräusch

Zahlenbereich

Computeranimation

Spezialrechner

Zufallszahlen

Computerspiel

Dichtefunktional

Digitale Photographie

Diskrete Simulation

Inhalt <Mathematik>

Abstand

Diskrepanz

Softwaretest

Approximation

Vektorgraphik

Systemaufruf

Ideal <Mathematik>

Elektronische Unterschrift

Approximation

Dichte <Physik>

Digitale Photographie

Rahmenproblem

Mereologie

Klumpenstichprobe

Standardabweichung

2:00:10

Videospiel

Web Site

Prozess <Physik>

Punkt

Rahmenproblem

Raum-Zeit

Mathematisierung

Relativitätstheorie

Voronoi-Diagramm

Gruppenkeim

Zellularer Automat

Geräusch

Extrempunkt

Raum-Zeit

Computeranimation

Diagramm

Datensatz

Rechter Winkel

Vorzeichen <Mathematik>

Ganze Funktion

Hypermedia

Kontrast <Statistik>

Merkmalsraum

Einflussgröße

Verschiebungsoperator

2:04:09

Subtraktion

Oval

Vektorgraphik

Rahmenproblem

Extrempunkt

Voronoi-Diagramm

Vektorraum

Zellularer Automat

Ähnlichkeitsgeometrie

Computeranimation

Generator <Informatik>

Rahmenproblem

Elektronische Unterschrift

Vorzeichen <Mathematik>

Gruppe <Mathematik>

Zellularer Automat

2:06:22

Punkt

Rahmenproblem

Klasse <Mathematik>

Programmverifikation

Voronoi-Diagramm

Vektorraum

Gesetz <Physik>

Rechenbuch

Raum-Zeit

Computeranimation

Spezifisches Volumen

Elektronische Unterschrift

Webforum

Diskrete Simulation

Schätzung

Abstand

Cluster <Rechnernetz>

Hilfesystem

Phasenumwandlung

Diskrete Wahrscheinlichkeitsverteilung

Videospiel

Vektorgraphik

Güte der Anpassung

Rechnen

Elektronische Unterschrift

Dialekt

Rahmenproblem

Einheit <Mathematik>

Basisvektor

2:09:52

Subtraktion

Punkt

Rahmenproblem

Extrempunkt

Hochdruck

Klasse <Mathematik>

Voronoi-Diagramm

Vektorraum

Zellularer Automat

Geräusch

Whiteboard

Computeranimation

Richtung

Kompakter Raum

Abstand

Cluster <Rechnernetz>

Folge <Mathematik>

Videospiel

Güte der Anpassung

Ähnlichkeitsgeometrie

Frequenz

Rahmenproblem

Zellularer Automat

Mereologie

Extreme programming

Klumpenstichprobe

Programmierumgebung

Lie-Gruppe

2:15:15

Subtraktion

Punkt

Konvexe Hülle

Klasse <Mathematik>

Kontextbezogenes System

Computeranimation

Persönliche Identifikationsnummer

Netzwerktopologie

Ungleichung

Ungleichung

Rechter Winkel

Kompakter Raum

Kompakter Raum

Dreieck

Abstand

Cluster <Rechnernetz>

Klumpenstichprobe

Beweistheorie

Grenzwertberechnung

2:21:13

Softwaretest

Folge <Mathematik>

Lineares Funktional

Subtraktion

Punkt

Rahmenproblem

Extrempunkt

Voronoi-Diagramm

Vektorraum

Elektronische Unterschrift

Whiteboard

Quick-Sort

Computeranimation

Softwaretest

Zufallszahlen

Elektronische Unterschrift

Zellularer Automat

Abstand

2:23:24

Lineares Funktional

Punkt

Voronoi-Diagramm

Computer

Ranking

Gesetz <Physik>

Dialekt

Computeranimation

Flächeninhalt

Rangstatistik

Gruppe <Mathematik>

Gerade

Fehlermeldung

2:25:22

Lineares Funktional

Vektorgraphik

Rahmenproblem

Gefrieren

Paarvergleich

IRIS-T

Ähnlichkeitsgeometrie

Physikalisches System

Computerunterstütztes Verfahren

Primideal

Ranking

Elektronische Unterschrift

Ranking

Ähnlichkeitsgeometrie

Teilbarkeit

Computeranimation

Eins

Rahmenproblem

Elektronische Unterschrift

Rangstatistik

Menge

Gruppe <Mathematik>

Mereologie

2:28:01

Information Retrieval

Punkt

Rahmenproblem

Leistungsbewertung

Klasse <Mathematik>

Partielle Differentiation

Zellularer Automat

Gesetz <Physik>

Ähnlichkeitsgeometrie

Verzerrungstensor

Computeranimation

Datenhaltung

Eins

Freiheitsgrad

Elektronische Unterschrift

Rangstatistik

Dichtefunktional

Spieltheorie

Diskrete Simulation

Schätzung

Minimum

Stichprobenumfang

Randomisierung

Information Retrieval

Folge <Mathematik>

Lineares Funktional

Zentrische Streckung

Vervollständigung <Mathematik>

Systemaufruf

Datenfluss

Elektronische Unterschrift

Entscheidungstheorie

Parkettierung

Rechter Winkel

Hypermedia

Verzerrungstensor

2:32:09

Information Retrieval

Resultante

Abstraktionsebene

Normalspannung

Computeranimation

### Metadaten

#### Formale Metadaten

Titel | Video Similarity (30.06.2011) |

Serientitel | Multimedia Databases |

Teil | 12 |

Anzahl der Teile | 14 |

Autor | Balke, Wolf-Tilo |

Mitwirkende | Homoceanu, Silviu |

Lizenz |
CC-Namensnennung - keine kommerzielle Nutzung 3.0 Deutschland: Sie dürfen das Werk bzw. den Inhalt zu jedem legalen und nicht-kommerziellen Zweck nutzen, verändern und in unveränderter oder veränderter Form vervielfältigen, verbreiten und öffentlich zugänglich machen, sofern Sie den Namen des Autors/Rechteinhabers in der von ihm festgelegten Weise nennen. |

DOI | 10.5446/345 |

Herausgeber | Technische Universität Braunschweig, Institut für Informationssysteme |

Erscheinungsjahr | 2011 |

Sprache | Englisch |

Produzent |
Technische Universität Braunschweig Institut für Informationssysteme Balke, Wolf-Tilo |

Produktionsjahr | 2011 |

Produktionsort | Braunschweig |

#### Inhaltliche Metadaten

Fachgebiet | Informatik |

Abstract | In this course, we examine the aspects regarding building multimedia database systems and give an insight into the used techniques. The course deals with content-specific retrieval of multimedia data. Basic issue is the efficient storage and subsequent retrieval of multimedia documents. The general structure of the course is: - Basic characteristics of multimedia databases - Evaluation of retrieval effectiveness, Precision-Recall Analysis - Semantic content of image-content search - Image representation, low-level and high-level features - Texture features, random-field models - Audio formats, sampling, metadata - Thematic search within music tracks - Query formulation in music databases - Media representation for video - Frame / Shot Detection, Event Detection - Video segmentation and video summarization - Video Indexing, MPEG-7 - Extraction of low-and high-level features - Integration of features and efficient similarity comparison - Indexing over inverted file index, indexing Gemini, R *- trees |