Add to Watchlist
Hidden Markov Model, Video Retrieval (09.06.11)
Series
Annotations
Transcript
00:00
All right but the body to today's lecture about multimedia databases so as it might has noticed and not of the bike and the and the 1st role is not to longtime that because the bike is currently attending a conference in Berlin and review but it too is remove yesterday and is now to speak that much at the moment there for a long period of time replacement sold on the road we are well trip to pay out the the paddock and the assault and the things we will be able to on Saulieu questions but the Disney questions to Canada and the and the offer to sue of the bike and they will definitely be a way to help you care about that start with a brief summary of the Plymouth lecture last week we are talking but grew by coming so that must be notified my guess on just saying a bit and discussing how all the possibility time what time walking works on the end of last week but also hit Mike of models have been introduced which also is a model for whom audio retriever rescued the ideas to use some kind of are of plant state machines to with to most the time development of music all right and that this is where we will continue to date so today after look at today's this is not a video a tree look so actually this long CIA so I continue with Michael Mols and the talk about the rich what it is and what is thought and and what you can do to make good beauty of the 1st part of as said about him my of of on this is my Pompey and Chris and will talk about the rich we've all right out as you might remember the might of model of essentially on some kind of state machines followed that they no just give example and try not to talk about for leaders that much the basic the idea is that you have state you want you to to 3 and maybe many many model and you have to and the distance between state so and each transition has a ability so and at any time of time 1 until maybe finishing time capital at every moment time we on exactly 1 of these states and state then changes of state of the Booker with the given abilities of each of these transition CIA has a publicity assigned to a sulphoxide though the probability of SIAM and at a given time in state to wonder then points on the probability of being in a state queue to move in the next time step is that says
03:27
Quanta and the probability of being in Kew read is read what so it's very likely if I'm now into 1 that I'm in the next next time step into 3 and not selected and through and the same is true for the UK the state of cause them might be loops half year might be a half ability and 2 3 might refer back to queue ones with Publity 1 this might be deterministic in some way so essentially as said is that the finance state machines no from theoretical along computer science of basic automaticity read but now it's a fall ballistic and but also important 0 as a Consignia with his abilities and the probability where you are in the next step off time only depends on the current state so it doesn't matter which way you talk in the in the in the Indian Michael to arrive at the current state Italy called to no currently I'm state take you to enter probabilities completed the determined the abilities where my would be in the next step so being at the scene the to these states and conditions in Michael more of also have observations so here that deal to have some possible observations with call them a 1 0 to without of called the might be many many more and in each state assumed that the state is and that by the machine with a grid some given ability 1 of the nation's is generated exactly 1 observation it and then touring state accused read at some given point in time with the probability of 20 per cent of the operation 0 to is made is generated by the state and the probability of point it's that saying that the 2nd observation is generated so observations might be out for example in order to achieve a found characteristics of features of the of the PM only a signal that we are able to absorb so usually when of an autistic as if he had large large of a stream of all 4 of sampled and became analyze assemble and make some kind of being from a string with with a sequence of events for some Blignaut load or something in between and this is what we can also what we do but we don't know how these observations had been generated by the underlying health of is what we don't know and the might of is used to model of all music is generated was seen in the last lecture that that the individual notes usually on divided into several steps so that overseas attack sustained became the model for the foot and these could be possible steps model by the underlying Michael processes and also variations on just something features became derived from the autistic 71 to find out what actually happening in the more alright is and finally they also of his distribution so usually and when you used by state automaton and as 1 state might as the the stopping states and had Michael as moving no single starting state but also the ability of science 0 if we stopped at a time at the 1st time step he was 1 than on their ability distribution Usually are usually described by variable called high and be determined but with which probability we stop in each of the possible states launched at the start distributed as a set of laws that has called was across that's not William Paulton to understand happening and just think of it as it as some kind of role ballistic Finite state machines are the operation states are the noted by Cuellar within the next observations under noted by small both so he area big and possible observations bombed the fiscally that it has music some of has already had already made 1 but that was more or more easy to understand the guest so with some initiative produced abuse and that is not shown in the picture so British which determines where we start and in the summer we have accepted the exact decayed sustained reuse and finish structure for the generation of individual notes of restocked somewhere Sodetz assuming that we start with the ability 1 in the state aid if we are at time time 1 in state aid and with the probability of a depressed and results of Asia's generated the ability of 20 per cent this observations generated the stops evasions can be generated when we and state aid so in the next over time we can days in state aid for the public to 50 per cent change to the state died with the public he of the deepest and Bulwara they well understood that to 1 side and link in transition and its 20 per cent going to date and so quite the and and said as soon as they enter a new state and observation is generated by the state carried alterations are abilities some so sort let's assuming we have given some of the nation's sequence review given some some UK of of which uses it is the noted by number and that the 2 we have
10:03
So exactly the observations which might be some features of the All You signalled we just received and re reuse autistic use of a read autistic not say what and then reuse some featured extractor and from board is signalled we arrive as sequence of the of the patients which tried to describe the autistic in some way so again this sequence follows the time so it might be the case that he had saved the violent integer assembled a combined into a single feature which simply might be some kind of of the patient type of says new make up Asian types 3 whatever that this might be it and the next 2 thousand sample to analyze we get future type 1 and so it goes on so we can describe of beaches cream in terms of the occupation of Michael model and let's assuming we have given some some states and some of this probability of Michael model that described in some way how his observations generated from each other for some example by using and and more like this which which include take delays sustained reuse for the generation of notes of Kaylyn now to some some coarticulation only because usually you have different different Michael Mols of available the described different things an autistic coach could do talks of all we have to be a with a mug of model that describes classical music how it is usually a generated and with a mug of model that describe pulp music and now we have some kind of observation streamed from from some piece of music and just want to know whether Paul a classical music so they have to to mock of of and and it would be interesting to know of full which of these 2 more looks the opposite the of the sequence that has been generated more like so it is more likely that the that classical in mock of small estimated self sufficient for the doubles the sequence and its most likely Jessica music otherwise of music so that could be a possible application of arrives what we want to compute the and is the probability that the random observations in the steps of some given did not go model is exactly the observation we made so the is allways random area will other T some of generated by the model of the small those or observations we did so these things up he arrived a came only to some some difficult relations and Andrew see how we how we can computed ability looking at the start of following again let's assuming that he is the sequence of Hughes is state sequence of some sequence of the state the machine went through and he steps and you want to compute the probability that we made the observation we made so there should be 0 1 equal smaller ones and the block and the quality is smaller Elodia so that you want to know the ability to make the observation we made given that the machine went through some specific sequence of states so actually this is quite he the because generations of the of the patients is completely independent of each other so it doesn't it doesn't change anything if if that we made in the previous state a certain observation as long as we we know the state in which we are so that all begins to independence here and write the probability as probability wives as a product of the club's ability and each by corresponds to each to add a single moment in time and each sector Bennett the probability that the make the observations we made at time came given that we we went through a fixed state sequence of 1 of the end of cause that as a single observation X time Kate doesn't depend on which states we have we have been in the possible will be in the future that they should publish but he's only depend on the state which could be are in than the considerably remove all these conditions and early keep OkCupid could small but cubic a few like a lot of the pages but clumsy sort of that but ideas that simply no in which take the I'd time K is the only information we need to so if you want to find out what the ability of making a given observations and we know that we went through a certain that the cost of states in the machine and the probability of this grew to product of automation probabilities corresponding to states we already know and of various slider think might give that reintroduced some sort 10 notation of the J K Isabeau ability of making observation came when we are in a state J also a IgE just the transition from ability to less makes some space here concede that a shot at notation a by Jade just need ability of moving from states Jane to state by of the prices of the more handyman describing or this from farright so this is just what or defined defined shorter and so the probability of possible hopes of observing observation of the Beijing Games given a certain state we are in the same go sorted notation of serving of the Beijing Games in the given state and all you need to know computer the product that arrives on the make and other computations some who want to know what the ability is of going through at a given sequence of states to its assuming we have where the sequence of the state and you want to know which way which with what ability we went through this serious of states and actually quite easy so Soweto we know that we started to think you J 1 and that this is a 1st step and all mockup rose as the probability of starting in this and the state is just the initial viability assigned to the state and then we just an hour or condition probabilities so you to the crude huge through the window at end of the 2nd stepped off time we on state he would J to and then we just have to at the ability for leaving state 1 and and touring data G J troops and this week to fall further steps under arrived at of finer and step of find state of mind so that this sort of publication of state sequences works and we also know that the following his through this due to the condition of the condition of ability to estimate remember condition of ability is defined as follows the ability of some events given another year and exactly this ability and those events happened at the same time divided by the ability that only become conditions event oppressed so if you premultiplied this wanted this we arrived at the home of easy for right but they found remember we want to know what deployability of making a given alterations is so we from UK of motivates observation and want to know with 0 with what ability to more generated this alteration has exactly what we are doing here so S we don't know which sequence of state we went through because we only know the observations because simply some up every possible sequence of of state so and as the sun disjointed and be indeed can some them up and then the probability for each of these sequences it's just a probability that you may cultivation given that I went through the exact might applied by the ability of making the last this year history patient this comes from a case of not being the previous light 3 with a computed how we can express these 2 0 abilities a different way and if you put all together and I arrived at the ability of making a given observation is the some or all possibilities sequences of states and each summoned contains the publicity off starting in the 1st state making the 1st observation and the state making a transition to a 2nd stage began making them the 2nd observation remade and so on August last Friday so the cost of their exponentially many different sequences year we have to try this just a falls approach which can work in practise and there for of them on grids available to compute this Publity inefficient way to come to that can do that to win that 1st let us led us to another computation on and is about the most probably state sequence we went through so again we a given some of the nation's sequence and we are a given some fixed did not of model so so I know the model we know the observation we made and the question is what is the state's sequence the model went through to generate self patient that has the highest coal ability to assembly of want to know which way the mock of low when through to its stated went to generate the observations and of cost as a principal in statistics so called maximum likelihood estimation and Maxine Michael estimation symplast as well if you have to to choose between different possibilities that not might have called on the event and take the most probable cause ability you have so when they do it to go to do now is to find out which is the most far away the most Pablo state sequence that generated off observation so the load almost this and we arrived at the following year you want to know the probability respect to are known model to go through state secrets given all observation and the heroes of the very able to we want to we want to find so that is what we know and this of area but and you want to maximise the Tote so how can we do this again so we want to know every week while I know note that this is true and this is what we want to maximise and we also know that the probability of making it all give observation this constant because on a sequence of observations is constant so we simply can forget about this term is just a dust constant and went on to maximise this but it would be the same as maximise this time because only of the cost infected changed him became so instead of maximising this condition probability of the 1st year the actually want to maximise we can also maximise this 1 we would get the same results but it is easy to compute or away again and the due this that defined some of the new area of function desert Shiite this some
23:46
But the probability of the most probably away to the end up in state Huai at times the given observation so we assuming that alterations constant we know what the also and that the idea of simply give us the probability of a maximum probability of arriving in state Hugh in time for T so that the definition so so we could computed by trying out all possible sequences up to time tea and just computing the all the fall of the of the ability of this sequence but has said this is quite a complicated so that it is a nice trick to do this and this is about making a because of the conditions of this function so found that density plus 1 J is the maximum probability of of arriving in state J at times she plus 1 so this is state James and we know that time he plus 1 we are in a state J and what is now the maximum probability of off of being there at the time he plus 1 so we can't computer cinemagoers a fashion as a sent the cause
25:25
We know when we are in a move that when we are in stage time he does 1 we must have the right staff from somewhere out so that many many other states have called on to 3 and a very many what we know now that we either came from the state 1 from state to of from state 3 of from some of the state will drawing up and of course that this is a terrific if it does people's 1 J is the maximum probability of arriving in stage at a time she plus 1 than of because this is the same as giving them if IFS's that and we know we arrived from state 1 and this month the this probability multiplied by the transition probability from state 1 to stay the game and as this is Maximo this is all this must be also maximum or other around if the baseball ability of cheap last 1 in state that they J is Maximo then this 1 also had most of the maximum for the state 1 so and then we can simply try out all possible states we might have come from my remember they accept the and states we might have arrived from and stage and then we take the maximum probability at the previous timestep time the demise ability and being in a state I'd had time team and multiplied with the condition by ability and just multiplied with the observation we made in time she plus 1 of state J and we just take place in maximum billion year because the maximum must belong to the to the to the States where we arrived from so that the music ability that something that the eye in this case might applied by the condition ability is the largest collective can get all possible path a retreat could have arrived at state J so this fiscal year because of the conditions of the time and the only 1 who want to compute is that the function of the written Achille we can use a technique all dynamic Programming which which is used by the so called the by tourism after Paul Epaulet being quite but all along
28:18
And that should not be too difficult to do so we start with the Delta in or initial state because this 1 is easy to easy to compute because the day before each state we do I find that I had time to use the maximum probability of being in a given state is simply the issue probability of stopping and the state's multiplied by the probability of making the observation we already know that the state of the function of a talk about soon which is defined to be 0 in the 1st for this and it is a step that we do not with and the and we are repeat secondstep for each point in time starting from 2 and ending by mighty T remedy is the last step in all a sequence of of the patients so and again for each state we now defined density of which is the maximum probability of being in state in time by using these this equations explain the period slides because as you might hermeticism I'd of scene we only needs to define density of something will indeed the desert minus 1 for this so are now hint inductively defined that functioned because we know the 1st step and and if he says he used to that only need them to the death toll wants and we can simply computer maximum over and state multiplied with this ability and and we can be fine or next disaster so and not this high functioned simply remember which way we talk when we entered the state J so you are taking a maximum over all possible ways to arrive in stage and the 2 5 functioned simply stalls which way when sodas UK Max this simply stalled for which state we came from the probability has been maximum so now we have stopped the maximum probabilities for each step in time for a full arriving at a given states and we know which way we came from for the writing this maximum probability so we had done after we of computed density and that that the big Seetso a 5th arrived on not 7 times of poetry observations and now we can simply read about how a probability of making observation and the maximum all of the staff at the time but she of because for each for each state that the T 1 of the key and states this again on the abilities of the of ending in state and at time T given that we made all made and the and the just take their the highest daily of these days because the corresponding state must be the state would finally ended in with the highest ability of this this year and now we know which state much what it was in this most likely sequence that has generated observation was last 8 this sequence again UK by Max here so the the idea that has the highest at the value last step is the last date in almost public in almost all the state sequence so now we need to reconstruct the past we went through the we window the last 8 and with before of high functioned we now can find out which have been the states that came before that so we know what that at times the team we have been in state J baked and when applying the functioning of charity that this will give us the most are the state became from when we know that the ended up in the States but the and so and by doing this fall for all time for all possibilities possible timesteps weekend reconstruct the path to band through from the back so now we know what is called what at the highest probability of generating observation and we know which is the cost postponing to the ability to which is the most likely pence we used to give this used to create this observation alright so it's a lot love complicated mass found an guessing you won't really understand all details of the takeover of it all much you if you think about it it is that difficult of the of the patients but clumsy and that
33:41
But just take a look at it and they did really really see the adjusted to combine some of the season thing about what really happened in the UK a number so what was camiknickers conclude not given a fixed it not of mobile and given some sequence of observations Ratawi underwritten by just presented gives the secrets of gives it as a sequence of state which has the highest ability of holding cultivation and the statistic as is a good terms the maximum nightly estimated terms fall far state secrets
34:24
Alright so not which can decide given observations which model cost it with with which the ability and the what has been happening in the smaller in the underlying staples so it also is often a problem and we where model to come from sold also I can start defining all I'm not of model in some way and hope that the the summer in some way that some kind of a new generation Croesus but usually what we have just sold divisions we know some some features and the kind of stuff in the music sequence but we don't expect the No 1 on the underlying mathematical Prosser scalding alterations so that deal with the 2 assuming that there is some did not of model of but we do not know what the transition probabilities are a probabilities are and what the observation abilities are these parameters has to be a have to be estimated from observation and some some training training sequence a chaotic ask then becomes to find out the parameters off some luck of model of giving the not so observation seat of long observation secret so we want to find the cause parameters all of that maximise the probability of of making all observation of cause the trading secret should not be to sort because usually it all are more less let's saying that we have in states and we need to estimated DataVision probabilities which is and time and we have to estimate of Division probabilities bridges and times they began with the number of observations and probability distribution is a vector with and components so that all we have to estimates roughly and scrap parameters but with which might be quite large and to do this of calls for the last training secretly lot of them to be able to estimated or parameters but of course we want to maximise the ability of the probability of generating observations and 2 because of so many parameters as quite squad public it highdemand Optimization problem but it can be speak and sold of Italy's approximately by the so called fallout algorithms on which ends up and local optimum so we will prove here that their books but give it idea what really down of the bombers were the 1st needs an initial estimates in his estimate of these but so some idea where the where we stop so sometimes we we have some additional about the amount of more led work here and we could simply defined all initially estimation of the starting abilities to be where the training said of which we know of which which states Machine went through we consume guesstimate account how often the a starting state has been state and hominy many training processes we have and it is simply the defection of what the same could be done for the parameters a and B so condition abilities and the ability of generating a given up the pace so just can't can can make use of various knowledge 3 half and make some a new should so if I have absolutely no idea we can justice you and in the code is to be using a unique on distribution for these parameters which also work and in this iterative across food Finite estimations under finely arrived at a local pub to move up Balkanisation from a case that iterative re estimation of the parameters this idea and to do this we need to defined to the area to types of area so that all the and that could various a where they called followed and that what we have at the very able to describe the past so we followed their of the of eyed and this is a probability that we make that you make the observations we made in the steps 1 2 seed and we are currently at a time she in state by so this is something about the past but had in the past let's assuming the and now it's time team we know that we are now in state eyed and other see eye is just the probability of arriving the of of being in a state by its time given that we made where the observations we actually made it to the right so that no some of calls we can we can be find offers of 1st step so if we are in time the 2 its 1 adult 1st step and time than the probability of being in state I'd is of calls the probability of stunning instead I'm multiplied by the ability of making the observation we made instead by entry can also make a weaker as a definition of all or or a although follow them he plus with because 1 so if we know of the team and we can be fined of keepers 1 of says of the probability of being in state Jade timekeeper's 1 which simply as the some or all of possible state we might have come from so a free on state J at time he plus 1 we must be in some other states that code by its time team and the probability of being in state idea time is the 40 of I'd multiplied by the transition probability of going from state by to state J and since the difference starting states all fully find disjointed and we just have to take some here and finishes finish this up as we know we are now in state J we just have to to multiply the pro ability of making the observations we made at a times she plus 1 given that we are now in a state of this way we can quite easy to find a serious all followed but very able for each step in time in each state the same can be done missile code Beckham Adam earlier this and that but not it very would do not speak about the cost that about the future of local so again led to see what we are not all in in that time died in some state I'd than that of the of finding is a pub ability of being ability that given that we are not all the time he in state by the probability of making the observations we made from their not so we gold some processes in the in the future people's 1 2 2 and so on and we don't know which way but given that we are not in state by what is gullibility of making it in the future the observations we may the actually made so again at the weekend I find this very able recursively of computer because of the because in the last day of time we know that if arrived at the end there are no future observations and the long so the probability of of making no observations when being in the last 8 of all Prosser is 1 because when they are in the last day of all posters and there are no observations services this is safe event which allways so we know the last 1 and now we can be defined the previous 1 so time died state by and we know PPL that from this state we must where we must go somewhere that in some state Chase at times she plus 1 and again we can make a list of the nation's summing up over all possibility states we could have gone to a state I'd might applying the condition probability in each case multiplying the probability of making the observation we made in the stage and multiplying the probability of the of making further observations from there is 1 with which so codes stage plus the next step in the future from the next step on from my guide to the future from next next on the definition of the future of the current time chloride only the ahead of in the next at this
44:34
Defining the poll ability of being in a state Huai in time the if we know that 0 has been of the soul and the followed Beckford very a we have assuming that we know we are in some states and you want to know the gullibility also making in the future of Asia's Olmec England made also patients in the past and now we want to know what the probability of being in some states that some given point in time this is a guy muchioved off eyed and is simply given observations while the probability of being at being in the state it's time to so this condition of probability you might remember by divination this is the 1st year event and the 2nd events occurred together divided by the ability of the 2nd event of this money simply is a definition of condition of ability And this ability is exactly this money because the probability of being in a state Huai in that time the and making all alterations we already made simply is a probability of the of the of of all followed by area will multiply Bulbeck various so now is cost multiplied by future after spoken of divided by the public of the nation And this can be expressed in this way just summing up for all possible states which could be time team the some kind of mobilisation so that the real now tell Sussex acting at the viability of being in a given state it a given time is this year we of seeing we weekend computer all fallen very in Rekers way which is why the weekend compute or Beckford very Ablett's quite easy way that for all weekend computer this gullibility also in the same way just summing up all matching followed that very able and Computing this racially fights to the next the ability is about transitioned in the society of the time so we know we want to know the probability of being in a a state Huai in time team and being estate huge that time he was 1 of time the state why and we know that the next step in time we are in stage to 1 of the probability of this piece of all state sequence happening zone of calls this does a definition here we know we have given all observation is no dozens of the probability of being in the States the and again we can just apply the definition of condition probability and we end up with this month so we know of the fat we arrived at stake by them must be a past the must have been some steps before led the probability for this happening can be expressed by of followed variable of some and the prop and there must be a past then them must be a future after we have been in stage and then we go into a lot of of other states PublicAffairs of that but very a bill that she was 1 of J soul and the only thing we need to connect here is a probability of making or observation NYTimes keepers 1 state James observation we already know again this divided by of Asian probability of this 1 year face the same and the division public can be expressed as some almost the combined combinations of light and shade of gave no cost points to the eyes and corresponds points to the to the Jay's we just some up the stone for all possible combinations of fine Chase and their phones Noman eyes of ability him so again we know how to compute that we know how to compute the a and B parameters of the model we have some estimation off and with this tool's as we can also compute white easy to review side for this transition to ability looking next 1 some some expected values that we might need so found what is expected the number of times state Huai was left of this simply to some or all or and them of area so we member of the amount that multiple ability of being that being in a state Huai time died and there for summing up for these abilities for possibilities the steps in time we have the expected value of times we have been in states state high so we know there is a sequence of states of laying the and on average hominy times we have been in state by the question this can express but the same can be can be done by the expected value of times we made transition from Huai to pick to huge this this some here and you might remember all go last estimated on possible parameters and also this condition abilities and if we had no more than the expected number of making condition for my 2 J that this high and also this ability must be high in all mode so of now the ball Majaida resolute does the following the uses only it definitions leaving just made and combined into and iterated estimation goals so we stopped so that we have and we have and across as the start the time on the with 2 0 0 and continues it should leave it iteratively just by increasing all by 1 each iteration so the the 1st integrations we find out what parameters come out exactly the way we find it before the end of the the area the sign exactly the way we find it before and of files and that can be computed from code estimation of the model parameters so we have had time but at step at a party Christian 0 we had initially estimation of all parameters we want estimates almost and and every step now weekend he realised postponing offers an that I've corresponding gum and size and from these estimations weekend estimated new values are parameters just by year taking the expected value of number of times we and at this stage and expected and number of transition we made from my 2 J so who died draws processes this step out secrets to 0 0 we have some initially estimates of by law and the use them to compute the offer that US gamma and size in this way and these again eye used to arrive in New estimation of their time apply for an and again we can be right and the other was that the car mine size from this and this continues under the has been some kind of convergence of all prose of these can be used to define the bees
53:49
And yes that it's quite intuitive to do that so so and this with all seen seen before this assembly simply be that the expected number of observations we made when not being in state by the end of serving Katie can be defined like this or a day at the ball but under the new law is by complicated but you can prove that they were uses this iterated roles as the following his true that the probability of generating also with off populist are plus 1 estimated model is higher than the previous estimation so every step of the bowlers who was all estimation gets better Lewontin or which parameters most likely generated off and and every step we become but closer to Iowa correct Solution of the most likely solution and at some point in time there is the a quality of almost equality and then the closest stops and we have estimated all parameters became a of the area from how can we can be applied not of most now expected to music reckoned recognition so as said the idea was that we have some of the stream of music works of the while file and from the stream of music teacher can be extracted so of the of the features giving some feature extract get a serious of features from the plight of some of probity of this body which featured essentially is a string of Peter type 3 fugitive 1 5th of 5 feature of something and so on and we also know that a series of features cost point because points to music of the event looks under the event underlying this scheme features of between and and this event is playing the No 2 seed for example of this could be a attack the sustained recent became the faulty to steps and it relies it is might correspond to hearing the No 2 seed on some from other high level event that humans can recognise so out again and of course how these features so that the teachers correspond to observation also of some did not got model we don't know how long legs and the looks like what state but we because we can use some training centre training data to estimated parameters of all I'm not of model and then find out how the see just been generated and lined the speech to the CBI
57:17
Arrived and what often does this time using different most different the band of and ideas of following we have an event called playing note the and the notes he is generated from some events weekend from some features we can so that they that had did that different ways to to generate this event so not notes might be generated by but this Features said but also might be generated with some by having along the sustained letters from the London decay level for something like this so that different ways that the cost of the event but at the human human noticed humans noticed another Dias to create a new not of model of for the event notes he such that the observations of the smaller other features a recognised and somewhere the back there is a Staples's in the muck of model we don't know something about the case so now uses data as follows we have given some set I just use of features 2nd series of features and you want to know whether the sequence of features coast wants to note see no died and let assuming we have a not of model of for note and in my of model of that describes how Load these generated then weekend computer what is the probability of making this observations with them but not of model for note seed and that the probability of making his observations using the mock of model for note died and let's say the probability here is stated press and the ability he is 5 per cent and I can be quite sure that the set of The just just of Sir cost point to the music event see of easy went the to went the built postponing not of model and then recently cheque if feels also some features some series of features which model is has most likely cost this sequence and events corresponding to mobile is the event that actually happened but their ideas of calls to to train my 2 billion not model for which the band a single and then determined which model of most probably called the of of the patient and the smaller the and hints at which began actually of of the these 2 to eventually went through admitted which might happen and if you have some time stream of music that I would would 1st divided into 2 steps extracted features from each segment and then try to find out which event was called the you not model estimation of an hour later you can use a ball outside a resident 1st to to create almost total part of this the of justice you that we have a model for each event availability and usually we do not know what the smallest actually look like someone other abilities but if we have some some under off what features but cultivations cost event a lataif large music database and some human manually that many earlier makes some some annotations Sam and told us as the event a could have been caused by the father before they receive sequence of features but it could can also be caused by this sequence of features and we have to each event and all the way to end the have long said that the large collection of different ways to create the event that they could use these sequences of observations to trained the parameters of what he not model this way we can arrived given some training data that it not that recognises or creates the events of time right so usually and 1 1 sticks to some some needs specialized more lives for example of the motives with their assuming that every possible conditions might happen is the most general Casey a called and got a global but here was the state's because the where estimated parameters refers to the side fall model of comedy state citing Tomaney operation types are there and then but and estimated Fault remain of abilities a one two for example and a 2 2 and so on we could also a strict also to a simpler model of where we assuming the beforehand that they are only transitions of this time so here as example of the power of the state and because you are a model of the probability of going from state to state 1 is allways 0 is ordered sound and when we know that this can make this can be used to simplify almost all estimation trousers because we do not have to estimated all possible parameters but only the parameters for those conditions that might actually occurred in order to see what it is about parameters and making all model but simpler and this is globally is last that we need to training by the idea again and and example of how this is done in the UK we have some some features sequence extracted from from off from off from all music for a piece of music this features cost wants to all of the patients may of by Michael Mols and now the idea is to find sequences in the long observation that cost point to specific it not 1 of to use the long series alterations that cause pointed different you vents and if in series of events that many many observations and but some of this might be the event notes this might be event no date and the problem is how to find these days of the find the positions where the event the band's change and that you want to find a sequence modelling exactly this and this can be expressed by using the by some in the and an for combining the founded detection of different the with events we can use so called macro hit mock of mobile which usually is an only not of most of what some you we have a not of model modeling the event notes he has been played here in Michael of model of modeling the event that note the has been played and he has Michael model for the No eat and we have Agunos observations here for example the event he has been as a coach you can't see as occurred even the again has occurred that the to combine the streets 3 into the macro and micro most where I can go from a model to model some kind of me time for macro model has called we stop somewhere so that is the musical ability would be to start with some event and now we can simply use this Makropulos to detect the what events mediocre and we don't need to tell my manually find out where the brakes between the events that to Assembly have large tomorrow and now we want to find out which is the most called the state sequence that generated alterations and we might find out that using the rhythms just showed that the most probable wave of going through this not promote when making this observation he is for example starting starting at the macro model died going around their somewhere in the States than making a shift to the scene model because he is a sea of and making some steps in the same are and then going back to the dealer and and from knowing the state sequence weekend we can conclude that the 1st 2 and must be the 2nd even must be seen the again must be a see so again the and the different types of macromolecule you look exactly the like like like a really good not most sewn up but now each state some kind of the that walks containing containing the completed not model so we can go from the macro model of from the from the technology the 1st Mellenby start at at at some stage in the 1st Load is effective way said so usually we do we really need to connect Pacific stop states with and states to if I was some kind of simply fighting mock of more of the for each year and then the usually is starting state and and ending state as an order attack sustain released a example of this might be attacked this might be the and states that the same could be true here is attack and this city and state in the only need to connect the and states to or starting state so it might be possible to go from event allied to avenge
1:08:33
By going from the ending state of model of all to the starting state of James of it might also happened that after after of serving the event and I'd again the idea event is of so that we need to go back to the starting states and the same is true here so when they get identified starting and and state we are able to connect individually not of model that that represent music events to and Ikramullah just by connecting starting and ending states are but the real problems of real Applications yet assets at usually is quite difficult to find out when a new and begins so what could be that tried to buy to find out subsequence this and it was again a sell off of the nation sequence
1:09:28
We could soon be used different piece of advice on the Ricoh stuff because the music mentation here we could also use a difference segmentations and use these things as an input to different Michael close to take which went Ritchie rent events happened at this point complicated because we have to have to find a lot of different segmentations and Load different length via many how many different observations the read we you want to treat at the single units And there is not much of what we can be used for that this is usually a quiet quite high complexity so Michael Mols on a quite general way of making making with the event detection in music the but they usually by costly but of the competition are so often that they are not then not useful for a week but it is tough on applications you have to determined the events that occurred initial amount of time but if you want to end up with large collections of music which could be done all said the and a lot of money on a good to would that can be used with care about what it used to ride window size for the subsequence does just sold so you that you choose W to be very high and that the large number of Michael of model can be traversed so and this is not propose and wipe out things through the different Model we can regular around 1 1 on the abilities for each for each way we went through and as soon as we detect a probability that is large enough and I can pick up the computation in a note that is that the starting sequence we of profound might be correct and then but can we can we can continue with a different we know that follows after that found that the 2 of that shows how it not of most can be used and that that was not prepared to do so a massive example year to you would do this and the next week and finally to a two see we load of occasions of and Michael loads of the sort the to about that the this we we had go Institute and time gold and this was cooperation with some up some researchers said that the fund and the reading of peaks so and if you if you want to raise pigs in and in the industrial fashion you usually have the problems that you want to want to know which takes and which ones are Sicily Alliance should be removed from the from the group to avoid the Multics get ill and the problem is that of calls of the people race picks could walk every every day and Innanen cheque all that they have if they are ill not but this might be because the pollsters because in the current times you to easily have thousand the 10th on the fixed that here that to race and the idea cities kind of automatic automatic and in this Peschici this was to be installed a microphone at the cage where they were picks up held and the simply listen to take out so the that and then find out for every car if the the biggest built 1 out of the question because we have not now we have and and in the event it took out and we have an event that has become might also happened and for each of these different types in the best diseases and not most places that describes this event and if not for the lack of a microphone we could not as the pick out which is represented by some sequence of 42 features then repeated the Saudi features to the 1st in the UK of model and to the the 2nd had not of model estimated probability that this observation has been caused by the use it or not the law load and ability that this of has been caused by the head of the had Michael mobile and if this probabilities higher than the cost was most likely from the opaque and if the cost and the ability for this in the UK of low has been higher because most likely be from and has taken if and when you take the big could be re could be removed from the group and being treated medical in a right to of the tests have been used by a basically the peaches used other similar to spectrogram features so you can slice the time into into small pieces and then each piece of has characteristic of the victim and his life can be combined into a single cost of for sold different different classes of slices had been the been I find some of this is time read and if so and if the sliced looks like this is not the time for something like this year is Yemeni many low frequency so that might be 10 for example and the world and this way we can begin to find a series of observations of different types of time slices of served and the and the stigma of actually this slices have been so small generally it has been observed that have usually as a duration of point for seconds and then months simply splits the time into full pieces starting time the main time and in some kind of decay time and for each of these events type has been assigned than a very simply not of model could be used so I'm audio some will follow the house
1:16:05
The sound is so that the new measures could body was found that the take look you move and try the qualifier itself about this physically all yet and the question is whether the stake is healthy the audio and giving the said Michael from all over the results have been quite successful so this they held it not of model of looks like it very easy to use and the different types of the events that might happen and that might become the combined in different ways and this is a not a model that has been trained for the for the sound of of different calf types of different diseases so you you could also instead of just estimating whether the biggest below not find out about which disease it has so we tried this and there is not much of an estimated parameters and the results of have been be quite good at think not accuracy off as occasion as being around 92 cents so this much much better than doing a visit inspection so when the 1st half happens the people race the the pigs and can the find out whether because not and has a nice at the case not model of don't have to be so complicated at this period at the usual even using in my of load are quite simple and results are quite good so next because they do it for you to down McChrystal should be Baker the break before that show them to make a break on Tilow 0 11 10
1:18:28
It sold after having all the fun of asylum picture and songs and pick up some of the things that are not continue for the next couple of weeks this video so of a digital video had been on area which opened with the broke on in the last like tenures plus minus across before of the of the young computer had not been refutable because you Tchibo more data but not days but you the handle the big bomb data Williams because if you try charges to show off times and Distributed Systems on all kind of things which are able to store transmit and even play the video and of course it is their usefully although that you can watch a video and detained but you can also to load of useful things to do like well presenting lectures media conference in having security videos and all kinds of the special the last couple of years in the area of internet entertainment leaders had been on a high rise special your own you your idea which which had been 1 of the fastest growing plotfinder from from the of fact that the aid of ideas and are contributed a lot of major traffic and Internet for quite some year's so he adds completely out dated flight from 2 thousand 6 unfortunately about where it was claimed that you trip across something like 7 millions of hours of video dead and you will from the duty free of that not days the amount of storage admitted it is absolutely this huge so I'm Powellite something like I'm several days of the Jim accurate which is Liu remove them all of computer ridiculous a here somebody claimed that the view that the country huge to 50 per cent of all of storage digital bombs because the extremely the bosses of the defined as really really large takes up a lot of memory that still a chose the scale of the retrievability storage casting of New York is very of of 1 of the major problems FuturePath his Scalability so that problem instead of millions of users and they all want to watch the continuously all the time from all over the world and they have really be problems with points and of some of them Hubert guy for sort of the window and the hazards of the school and the only 5 million people want to watch this guy falling out of a window the idea after the stringent all countries also the and that the single biggest very very very very challenging and and if you're really into how it is done on the actress sites or better based side of the future then you can come to election next domestic which distributed data storage very interesting and there would you hawkish Dorothy huge amounts of debt that in this lecture will deal with well they do so not just started inch of its and by but dealing with declined to sort out the problems we have with the year that we have to storage now distributed data storage problem you need huge on the beach basis may be great erased may be survey found depends on how much you have to make it accessible so to a we have something like Codex and you have to have Sorowitsch retrieval techniques for example something like Eskew L on it or information which techniques and you have to find the correct but it is still looking for like the assault searched and function for the most that have been today Baume to do the job quite development and to the so and and this year they have from just tabloid data of everything but should into tabled which are and that is chorus and a as soon as staff does not fit into a table it won't broke out quite about with almost that this is just the point lapses or just huge binary objects of a lack of food in your baby my will also have along with that decided and you use your overlooking the relation database you stocked with using match data for a tree so you have a video of summer are made in a book and then you have some columns which described it as high as well suited to the length of the EU's code a resolution would ever arrested the and then you can ask of became the from your would be used the name and the resolution and return to the all ever does not respect actually your country which is quite tricky them through the next thing which is done quite often this San splitting the review into keyframe just take civilians but is just before few the end of pictures split into key keyframe and then do the simple image based reputedly to key for that
1:24:16
Not that many techniques in for fashion rate which you read this you which we thought was a had been unlocked of tries and trains to get and media techniques into relations and 1 company or tried allows them as they are Ms or during the media or extend order and the base and yet more through combined you but to which is a complete tradition of classical for relations database 3 0 school of thought this year capabilities an audio and image and all the things multimedia extension And this extension well Faustus continued to develop a new more and that it was not also included into the protect itself as a more than banished from its very sad thing but there are a couple of reasons for that on the 1 hand and I'd ever happen very good approach for multi richly so that they said he was completely sticking to the relation along the keeping with all the table and then having a huge step of stop procedures and you define functions and some utility programmes which takes a multimedia collection for the year and images audio and then Jewish out some data and put them into a table and just below the multimedia objects themselves also the relations and then you can have Esquel to you read some of the at attributer and return the object that by this idea was quite nice to star is a mix of the of the destruction of the 1 hand it's very slow told the unextracted rose and stop with which have not been beaten High Performance and the the hand is probably not for the need for for example that if you really have a huge the you bomb you publicly can storage in relation database of fact if you choose system over the 2nd half to be troops are in the middle and then just put you can trillion videos into the studio to your own for were because relations databases and not really boat for high Scalability and as long as you have a duty to date Ivanovic road tremendously and you don't need high Scalability and high speeds of evidence is not really what relationship are about but it comes down to images this might have brought quite well part of the Imagine itself having an on line shoppe for something have all the shopping information and and all the pictures of Sarah in an integrated approaches relations database and the images which were quite well but still the whole complete multimedia package some all went down because it was not exactly but industry needed so for this kind of occasion have different kind of databases or load of no scale databases which used to transport the times are more Scalable and have different Ritchie return that you can't do not realise just 1 hour before the shock of look into what multimedia extend what and then you some of the area for later databases and multimedia databases themselves to match really rely on relations Caballo data but have different which in the case will be the top of the next couple of weeks the other non Esquel techniques for dealing with all the vigour the a sending off and most of Kate half the extent of broken under said on Friday and then you integrate or you much immediately into ASCO al and you'll know as guiding or drugs and now you should have some extensions in scale for achieving images of the so you have to go to your if you don't have that you do you have 1 table with all the matter that and then you are media 5 of them or not frightened you need to treat 1 for getting thematic data and 1 for tree and a multimedia point said it had some sort point for actually playing the clip and converting and so that was a new play included some kind of the conversion from whom for much to the next and then put it a was also providing support for which leaving some additional method data which was not manually and sort of the for for example I was able troops broader than really impressive but like when most media added to their banque or about which plans and had some additional features for extracting stuff like frame related to or colours Alooshe from tourist Lucian Melling's ought to just dumped in a year and that it expected its basic information storey addition cable and then you could accessible sing using and it was all built 1 you define functions and users to find features of beaches and the most the biggest selling point was having a unified ecosystem that all and 1 central database and of cause and status and once and for the nation that is your all the benefits like humiliating and especially recovered security rehab axes control have sustained user management new half bag of strategies have some kind of trans actuall some benefits that this was the main by the UK a as or a told the whole thing failed rail moralists bomb as far as no commercial real who does really gets a part of my immediate which uses or a class on the part in such about
1:30:42
All the while we suggest as take away at multimedia especially video inflation databases not mix fell had been tried not to renew that suppurate ago and Tyler the does not mean that multimedia ordered your databases do not to get just brought the your relations database of what music is different kind of database is different Europe are different dreary engine fashion different storage paradigms like all this cloth storey things they are quite get for starring media and now we need to retrieve techniques retrieval techniques this is a puppet of the selection there was a problem as a video of the video is unfortunately a continuous media means you half time and in time of the year before process and for every time I have a lot of fish from Asia so fabric timestep you have a picture of your frame which picture you could treated like the picture 3 after a couple of weeks ago in a mantra Tree semi have long off bomb or new music speech Road different languages like on TV leverage language German of French ever you could have your tax stockpiles or additional summarised formation and picture Falmouth and this summary of all this context that makes over you and you must act is of cost base more difficult and text or just an image across now we have it in for a new and mixed and computing the which meant
1:32:54
The rule of a big problem is that you can look at the give point of time out of the isolated of cause account but you want to available to read using the if you just have isolated 1 is England free of what actually overlooking the literature you have to look on the whole would set off a framed through the continuous M space for the whole thing should be treated as a document and not a collection of single the images for example and the food is a kind of cost to the same things that you have done this music you can abstract the music's a basic information lost about my out for example of the way from recording and the of cause you could do the future expected have some high higher level features like the music of both legs you do you F fought ever and then you could might tried to detect a melody of can be done with the so he have sustained basic information and made for the audience track like because of for the images and then you get a hell of a lot of structure of image itself to order the itself out and now we can detect additional features in the defeat for example something like a cut of the camera which of and judges on again something like that happens camera moved around them something like that some interested in a schmoozed detecting explosion these kinds of things Biofutures which can be detected also so that kind of queues could have missed deal well I'm a for the user can have really much acute a year and they ought not ask you all that is very funny and very I'm well unique a something which works quite well but of multimedia arias is period by example so for example you could have really like a Kate give me all of it is this Act or just show a picture place can you could stand to all the and about try to find that actor and of cause you won't be able to do a direct match of of the picture this year because the ball and and less as cream shot you won't find the exact picture in the media and the public in a different and annual different light named may be some under no different classes fought and you could you have using that complex here by examples of the type which use area of beauty and a lot of stuff going to semantics something like the of investment research at aguacate iLike's this movie died a move which is that this will not even more complex new you need more information for finding something fitting to that I'm here period of a
1:36:08
Yet still which troops to the country in the sexually have some basic information on a mission to the recovery of the images that some level future have of features and a half old algorithms available for dealing with images said on behalf of a man arrested for your trilled some like and Michael than detecting pictures and things like that bomb the can do what you could measures and and following is on track Sammy after the text retrieval and think this lecture but denied for making the trip to the judge on the show and have from the transcript summarisation and searching tax and using all of the twisting or can build something which could brokenby so assuming they have the biggest him available you streams and maybe subtitled has placate of things you can go out for a man just focusing on just stream SAP couple of basic things you can do about on the 1 hand you can try to detect the and the basic and is 1 c is somehow close semantic unit so if cameraman decided to cut by the US for the reason for the because the next thing the show something else of detecting seems quite important under yoga load of information Then you could try to detect person of staff in the US or in general so you could try to detect a key beaches back from 2 the just walk round from objects which are for crimes beach cars which she the rich people so maybe you could try to identify the for for subject to want to try to build a movie iPod and you want to know which seems contains which actress that might be interesting on a few more into security and have a new protracted do something like a cake you have observation camera maybe you don't want and equipped to fight the and attack and a contract by the people by and then try to detect by which is not common for for example of some of the time the number of people using freezing tremendously advancing has happened publicly and this kind of thing that you could try to detect voices of actress and then through some kind of assignment Cahill that was the problem the actress in this time of year and publicist voice belongs to the tree so you could be classified as a samples of detects puture or detective what you suffered some after a football game and you are not really interested in football but you want to breakfast a colleague's next day that she walked a match also yet not really interested at to could try to do with the kind of automatic generator summarisation off just impact part to addressing the just going like 85 minutes out and exciting 5 minutes from time are still contained in the and you could detect the things on a different characteristics like got people yelling or of movement on the play field which might indicates something interesting happening on soft and of course the Tampa Bay Area of things really really important so it's not enough to detect on image that there is a person on a match that at some point before but you to detect booking as a person as a person move has a car and move by from left arrived from right to left and a different and of different speed and this from a shorter can be really useful for use cases that especially the security communities very very keen on detecting movements and detecting time suspicious movement probably know this is a felt Rojek along the that they have security cameras like every day and a half like some 500 security camera Sheets's at which Allco into 1 control room and the beginnings after a couple of guys sitting there and staring at the of all the time in trying to detect some kind of cause I is something happened I was watching the of the screen and the by undetected said recently I'm as their worst night he of crops also time so you just take all the biggest change and Distributed through huge numbers of and cheese and people adjust such as acuity in their free time and was something suspicious and the just taxability and then and but through security expert who use a video and that some of its delayed but still you have acute picture of well some people died order of causes not would be tried to do this automatic annualised all this should have more automatically so we want to attacked the sums somebody steps into a restricted Area if there is a broad I'm starting at the street and you can detect the if we are able to detect a new object to the detect what objects are more less them from the images and then detect a whole they moved and he via the huge advantage over just playing image analyses because they do I have much to pull shot of the same scene and you your not limited of detecting say shape of object to identify but you can also also of its movement and that if has something moving around the city no it's probably not the house because houses usually standalone for the ball to be a car or a personal on any more sums up the mood of cost during moving up checked can also be exploited and different a area of under when it comes to compressing video of the year but you itself is really big uses a lot of memory but when you come prosecute and put it into a manageable area and 1 of the earliest ideas for compressing but it was quite straightforward Naughtiest dream and you have the huge number of images and you just compressed imagery and you compressed audio and then just put something country while you could do that would be not very efficient than the happy or fistic aided ideas of just compressing every 10 image and then so quick keyframe so every can image for example is compressed completely installed completely and then for the next images set just the difference between the current emerged in the last keyframe just have something and you that idea February last year and that if you have a video after all Cronje so persons and staff moving around a new back row have cause back from space more than the same in each unethical momentous hanging around like crazy you background is more on the table by the for on movie and this observation had for example being exploited by the new owner of the deal code 0 on all as using starring and bake for and David standard to separated walk round from the council moving up from moving object and then you could compressed the back from and have this on the back on which they are usually I'm just tiny amount of information needed for that same compressed and then have some just moving trajectories just all the for front objects are moving around and this idea was really great tricks to the issue remember older digital media for most like and to which is used for example the Neediest which takes up like a for gigabyte for just need equal quality and you could achieve more necessary 22 using from all more than 2 6 coated with just 1 day of for example as may need you well more clever encoding and for example chicks like separating for on the island and the cost of some Hollaback fight at 1 point was quite interesting and though their optimizing that stuff for them as a global championship because they are not a lot of digital movie trip and was supposed to be sold to people and the technology heavy ready for that and of cost should think of thought that came half a lot of static back complex or cream process which doesn't move or semi of players are well on TV screens that it and then you have a bad just you like tiny and has almost moving the players are moving and their completed as moving objects attack that the only algorithms and then compressed and across compression regular too big some of the Bulls disappearing because which just tiny moving up to the edge of players with all the ball from the book a thoroughly the care as it is necessary to make it work and that you will not now board and not invisible and football games
1:46:02
The basic cost the gas for next week detecting offcamera movement and he is a basic is camera movement is some all indicating that something interesting cut probably indicates Museum in your of indicates that the new Quentin unit is starting to reuse ideas about writing some semantic so that I shall semantic information from the futures of images over time a as stuff Flextreme meanwhile panning indicates and now Trucost shift from something different and Amurdag guests that detecting scenes and movement and cameras that will be on the of next lecture Sam the have a lot of other features like trajectory subject a hammered by detecting trajectory the big problem is identifying and object which is moving and then finding Howard is moving out of the examples image year versus gold was flying and you have much in the frame of the goal by now we can see that the goal flying you and and having a good threedimensional levels of Gulf caused can you could detect the has bullfight I'm especially the identifying trajectories is also very interesting inclined fighting for example that there have large I'm the you material some well happening and then you try to analyze what was happening for 1st time is that a car exit and and you want to know if you try boss breaking about below enough or if he did not pay exchange tensions then you could try to extract trajectory off the car and the other Communism and assembled and can extract automatically from the video then which objectless moving house price into the air and then you can the basic people behaving correctly breaking quickly trying to invade each other or not and this can be done quite efficiently and automatic the her back to the UK she suffered some before you choose if you try to find the during UKeU and there is some kind of magic retrieval at which just returned the best the your objects by what other means set of cost you have to have a recite presentation your owner that from Google by some of the and 3 into the Google texts such and then you get just the resultless the headlines and racial summary but however the book was a video are more Kaufhaus you coach as a display the name of the video which was off not really useful and somehow you have to accused the user shot summary of what is happening here and also you really have a lot of time techniques available which she discusses the next couple of weeks and more than the 10 some simple summarisation techniques like just taking the 1st 2 frames which is used black and doesn't work quite well and protosuchus take a random frame but you could also try to extract the most interesting frame of the frames using contains the most information of cost most interesting frame can be recognised by allies in the hope of finding frames which are dealt characteristic for the same you could throughout shot summary slacks lightshows having much in the frame after 1 after the other and 1 of the most difficult things which it could use it as a real semantic summarisation off best some kind of rising the something like trade for example of the movie which is for 2 hours you have a trailer which is like 1 hour which tried on away what the movie is about and what names and them for example automatic trader generational automated the summarisation also very interesting research public because you just can't not Ekstract the key frames and just put them together so it needs to be still a flew to view which is pleasant to watch for for example cup of his again time I've seen them off to research for the attack which took a movie trailers for action movies that specialized action movies as a base idea was looking for Load seek fences and fans and pudding some kind of various music that behind at so far by the and then looking for all the keen the lot of playing happens for what of yellow card which expounds quickly and and just taking all the things with a lot of explosions and then taking off as a lot of fights and a lot of guns and just put in some altogether and then I have a trade off making more of and at it looked quite well so on my Stephen automatically generated trade terminated to and it was not really role at Tuesday but quite good so this approaches converse and terms you not for traders but for the search for what it is a good idea for some a raising the intellectual clips set as can you do this study retrieve of cross and Attainment entertainment is huge sector quite now because of a lot of money to make a attainments Bansal load arias here so for example you and amount the new for the future of stuff is your clips bomber then you could have won using faxes you machines richest across television screen from from broadcast said could have something like some appetising to catch on the long time ago there were side you on to Santa the Flat in the data program such Michael Caine advertisment starting all advertisments that ending up and of classes and through the the 1st year across the pit on the map that were exploded that future and removed by advertisments from the data recording of classes means less affected advertisments and loss of money for the broadcasting companies of the flag was removed and now well you can detect that reduce easily any more and you need additional algorithms who can still separated for example content from advertisments off classifying use for ever in the UK is on for the last like a bit an additional very important cations security those a curated community in out among the payment security probably 1 of the community which does the most research in this area passionate comes of extracting losses trajectories things and detecting of movement pad as the identifying objects as or security staff and entertaining and payment of detecting scenes detecting content and by the 14th things out circles and this lecture I have to continue last lecture with a lot of money and just try to remember that much of more that looked disgusting and complicated but there are quite easy and even easier mock of modest you can detect quiet whose and just remember succussion picks with probably every days much occasion that you can do with the fact that I'm probably you will find an area in the late at the could fly map of techniques and all your a tree with 2 presenting which were reduced which people was like and
1:54:34
However you choose which was founded with the shot Overview and the real and the next week and there are probably dived into into all the things detection away from next coup like shut detect sugar and media stretch the abstractions like going from a really basically during from nation like single framed single Burton's timeless and a small part of what it used to be the stuff like both or detecting explusion are taking a fighting seen would ever found the topic of the next few said things for today and the fundamentals of from the view of some later
00:00
Point (geometry)
Query language
Group action
Musical ensemble
State of matter
Multiplication sign
Scientific modelling
Distance
Information retrieval
Frequency
Queue (abstract data type)
Multimedia
Aerodynamics
Software developer
Moment (mathematics)
Scientific modelling
Iterated function system
Bit
Markov chain
Computer animation
Network topology
Database
Information retrieval
Phase transition
Multimedia
Finitestate machine
Code
Representation (politics)
01:29
State observer
Query language
Musical ensemble
Group action
State of matter
Multiplication sign
Scientific modelling
1 (number)
Information retrieval
Invariant (mathematics)
Mathematics
Linker (computing)
Process (computing)
Area
Electric generator
Process (computing)
Structural load
Scientific modelling
Automaton
Hidden Markov model
Sequence
Demoscene
Flow separation
Order (biology)
Computer science
Code
Representation (politics)
Bounded variation
Reading (process)
Resultant
Ocean current
Point (geometry)
Set (mathematics)
Distribution (mathematics)
Characteristic polynomial
Virtual machine
Streaming media
Event horizon
Causality
Operator (mathematics)
Queue (abstract data type)
Data structure
Condition number
Physical law
State of matter
Set (mathematics)
System call
Markov chain
Loop (music)
Computer animation
Finitestate machine
07:44
Logical constant
Principal ideal
State observer
Group action
Musical ensemble
State of matter
Model theory
Scientific modelling
Multiplication sign
Correspondence (mathematics)
1 (number)
Mathematics
Positional notation
Singleprecision floatingpoint format
Process (computing)
Area
Spacetime
Product (category theory)
Electric generator
Theory of relativity
Block (periodic table)
Structural load
Moment (mathematics)
Scientific modelling
Sampling (statistics)
Hidden Markov model
Sequence
Functional (mathematics)
Maximum likelihood
Right angle
Whiteboard
Quicksort
Data type
Reading (process)
Resultant
Web page
Statistics
Set (mathematics)
Virtual machine
Computer
Event horizon
Sequence
Latent heat
Performance appraisal
Causality
Term (mathematics)
Integer
Subtraction
Conditional probability
Condition number
Data type
Information
Computer
Distribution (mathematics)
State of matter
Independence (probability theory)
Cartesian coordinate system
Markov chain
Estimator
Stochastic
Computer animation
Personal digital assistant
Game theory
Window
11:19
Finite element method
Performance appraisal
Computer animation
Model theory
Hidden Markov model
Data type
Sequence
12:55
Performance appraisal
Computer animation
Model theory
State of matter
Hidden Markov model
Sequence
15:40
Set (mathematics)
Distribution (mathematics)
Model theory
Scientific modelling
State of matter
Hidden Markov model
Total S.A.
Markov chain
Invariant (mathematics)
Performance appraisal
Stochastic
Computer animation
Process (computing)
Units of measurement
Data type
20:48
Performance appraisal
Computer animation
State of matter
Hidden Markov model
Estimator
Sequence
Maxima and minima
23:20
State observer
Computer programming
Musical ensemble
State of matter
Multiplication sign
Markov chain
Iterated function system
Average
Computer
Estimator
Sequence
Maxima and minima
Performance appraisal
Population density
Causality
Aerodynamics
Condition number
State of matter
Staff (military)
Computer programming
Functional (mathematics)
Sequence
Maxima and minima
Computer animation
Personal digital assistant
5 (number)
28:07
Point (geometry)
State observer
Slide rule
Algorithm
State of matter
Set (mathematics)
Multiplication sign
Mass
Computer
Sequence
Maxima and minima
Frequency
Performance appraisal
Population density
Raw image format
Key (cryptography)
State of matter
Staff (military)
Functional (mathematics)
Sequence
Demoscene
Maxima and minima
Computer animation
Equation
ViterbiAlgorithmus
Window
Writing
29:44
State observer
Statistics
Algorithm
State of matter
Set (mathematics)
State of matter
Hidden Markov model
Sequence
Estimator
Number
Maxima and minima
Sequence
Maxima and minima
Performance appraisal
Radical (chemistry)
Computer animation
Term (mathematics)
System identification
ViterbiAlgorithmus
Recursion
34:24
State observer
Musical ensemble
Code
State of matter
Multiplication sign
Scientific modelling
Markov chain
Parameter (computer programming)
Mathematics
Local ring
Area
Algorithm
Process (computing)
Electric generator
Optimization problem
Scientific modelling
Electronic mailing list
Parameter (computer programming)
Hidden Markov model
Sequence
Inflection point
Vector space
Quicksort
Data type
Task (computing)
Ocean current
Finitismus
Service (economics)
Algorithm
Connectivity (graph theory)
Distribution (mathematics)
Electronic program guide
Virtual machine
Event horizon
Code
Wave packet
Number
Sequence
Causality
Bridging (networking)
Conditional probability
Subtraction
Condition number
Division (mathematics)
System call
Estimator
Uniform boundedness principle
Probability distribution
Computer animation
Personal digital assistant
Iteration
Mathematical optimization
35:57
Computer animation
Algorithm
Estimation
Set (mathematics)
Scientific modelling
State of matter
Parameter (computer programming)
Mathematical optimization
Local ring
Task (computing)
Sequence
Maxima and minima
38:25
Point (geometry)
State observer
Group action
Computer file
Algorithm
State of matter
INTEGRAL
Code
Correspondence (mathematics)
Multiplication sign
Scientific modelling
Combinational logic
Parameter (computer programming)
8 (number)
Computer
Event horizon
Variable (mathematics)
Number
Sign (mathematics)
Iteration
Statistics
Conditional probability
Condition number
Scalable Coherent Interface
Area
Multiplication sign
Time zone
Electronic data interchange
Process (computing)
LTI system theory
Computer
Gamma function
Physical law
State of matter
Parameter (computer programming)
Division (mathematics)
Variable (mathematics)
System call
Sequence
Estimator
Maxima and minima
Data mining
Number
Computer animation
Estimation
Iteration
Asynchronous Transfer Mode
53:44
Point (geometry)
State observer
Musical ensemble
Numbering scheme
Computer file
State of matter
Model theory
Equaliser (mathematics)
Multiplication sign
Scientific modelling
Parameter (computer programming)
Streaming media
Event horizon
Number
Wave packet
Sequence
String (computer science)
Energy level
Area
Series (mathematics)
Pattern recognition
Physical law
State of matter
Parameter (computer programming)
Hidden Markov model
Estimator
Maxima and minima
Frequency
Computer animation
Estimation
Speech synthesis
Data type
55:12
Pattern recognition
Model theory
State of matter
Scientific modelling
Merkmalsextraktion
Hidden Markov model
Estimator
Sequence
Maxima and minima
Event horizon
Computer animation
String (computer science)
Implementation
56:48
Point (geometry)
Pattern recognition
State observer
Musical ensemble
Group action
State of matter
Model theory
Scientific modelling
Multiplication sign
Merkmalsextraktion
Streaming media
Parameter (computer programming)
Mereology
Event horizon
Computer
Estimator
Wave packet
Power (physics)
Sequence
Mathematics
Operator (mathematics)
Database
Energy level
Implementation
Macro (computer science)
Subtraction
Position operator
Condition number
Series (mathematics)
Shift operator
Structural load
State of matter
Scientific modelling
Set (mathematics)
Hidden Markov model
Sequence
System call
Demoscene
Estimator
Connected space
Maxima and minima
Event horizon
Computer animation
Personal digital assistant
String (computer science)
Order (biology)
Data type
59:59
Event horizon
Computer animation
Texture mapping
Attribute grammar
Hidden Markov model
Wave packet
Geometric quantization
Sequence
1:02:09
Model theory
State of matter
Scientific modelling
Length
Merkmalsextraktion
Complete metric space
Hidden Markov model
Sequence
Inclusion map
Computer animation
Ergodentheorie
Right angle
Library (computing)
1:05:02
Macro (computer science)
Event horizon
Computer animation
Algorithm
Scientific modelling
State of matter
Graph (mathematics)
Hidden Markov model
ViterbiAlgorithmus
Sequence
1:07:17
Musical ensemble
State of matter
Graph (mathematics)
Real number
Scientific modelling
State of matter
Hidden Markov model
Cartesian coordinate system
Sequence
Event horizon
Singleprecision floatingpoint format
Macro (computer science)
Computer animation
Graph (mathematics)
1:09:04
State observer
Complex (psychology)
Musical ensemble
Length
Scientific modelling
Multiplication sign
Abklingzeit
1 (number)
Disk readandwrite head
Video game
Moving average
Social class
Building
Real number
Structural load
Point (geometry)
Hidden Markov model
Sequence
Macro (computer science)
output
Right angle
Quicksort
Data type
Reading (process)
Point (geometry)
Algorithm
Characteristic polynomial
Mass
Event horizon
Number
2 (number)
Sequence
Program slicing
Data stream
Software testing
Subtraction
Units of measurement
Window
Series (mathematics)
Computer
Physical law
State of matter
Cartesian coordinate system
System call
Local Group
Number
Event horizon
Computer animation
Kolmogorov complexity
Graph (mathematics)
ViterbiAlgorithmus
Window
1:11:49
Scientific modelling
Structural load
Control flow
Parameter (computer programming)
Hidden Markov model
Event horizon
Measurement
Route of administration
Frequency
Frequency
Computer animation
Personal digital assistant
Uniform resource name
Data type
Subtraction
Resultant
1:16:36
Type theory
Computer animation
Algorithm
Physical law
Hidden Markov model
ViterbiAlgorithmus
Discrete element method
Wave packet
1:18:10
Length
Multiplication sign
View (database)
Archaeological field survey
Information retrieval
Heegaard splitting
Casting (performing arts)
Type theory
Hypermedia
Information
Information security
Area
Boss Corporation
Relational database
Structural load
Software developer
Basis (linear algebra)
Staff (military)
Hidden Markov model
Storage area network
Functional (mathematics)
Flow separation
Process (computing)
Internetworking
Network topology
Data storage device
Website
Quicksort
Simulation
Digitizing
Point (geometry)
Readonly memory
Algorithm
Discrete element method
Computer
Scalability
Internetworking
Scaling (geometry)
Hazard (2005 film)
Key (cryptography)
Information
Wave packet
Table (information)
Computer animation
Digitale Videotechnik
Information retrieval
Object (grammar)
ViterbiAlgorithmus
YouTube
Window
Matching (graph theory)
1:19:55
Data management
Information retrieval
Frame problem
Number
Computer animation
Smart card
Volume (thermodynamics)
Digital signal
Scalability
1:24:13
Query language
Multiplication sign
Complete metric space
Mereology
Information retrieval
Programmer (hardware)
Type theory
Bit rate
Strategy game
Veryhighbitrate digital subscriber line
Hypermedia
Multimedia
Control theory
Information
Data conversion
Website
Extension (kinesiology)
Information security
Physical system
Social class
Injektivität
Area
Theory of relativity
Relational database
Data recovery
Structural load
Attribute grammar
Functional (mathematics)
Latent heat
Network topology
Database
Data storage device
Order (biology)
Procedural programming
Quicksort
Information security
Prototype
Data management
Classical physics
Point (geometry)
Real number
Electronic mailing list
Scalability
Wave packet
Goodness of fit
Centralizer and normalizer
Causality
Database
Utility software
Subtraction
Addition
Multiplication
Scaling (geometry)
Information
Planning
Coma Berenices
Client (computing)
Letterpress printing
Line (geometry)
Multilateration
Power (physics)
Frame problem
Table (information)
Singleprecision floatingpoint format
Broadcasting (networking)
Computer animation
Personal digital assistant
Function (mathematics)
Mixed reality
Multimedia
Object (grammar)
1:30:20
Query language
Musical ensemble
Context awareness
Multiplication sign
Computergenerated imagery
File format
Visual system
Angle
Demoscene
Formal language
Information retrieval
Hypermedia
Database
Selectivity (electronic)
Multimedia
Information
Subtraction
Programming paradigm
Process (computing)
Variety (linguistics)
Continuous track
View (database)
File format
Relational database
Server (computing)
Flash memory
Computer file
Attribute grammar
Streaming media
Web browser
Frame problem
Number
Computer animation
Network topology
Database
Data storage device
Information retrieval
Mixed reality
Speech synthesis
1:32:53
Point (geometry)
Frame problem
Spacetime
Query language
Musical ensemble
Multiplication sign
Computergenerated imagery
Visual system
Content (media)
Semantics (computer science)
Rule of inference
Table (information)
Information retrieval
Frequency
Causality
Hypermedia
Queue (abstract data type)
Energy level
Multimedia
Data structure
Analytic continuation
Abstraction
Social class
Selforganization
Area
Spacetime
Information
Mereology
Content (media)
Sample (statistics)
Computer animation
Freeware
Film editing
Data type
Row (database)
1:36:07
Pattern recognition
State observer
Spacetime
Digital media
Code
Multiplication sign
Design by contract
Shape (magazine)
Mereology
Demoscene
Independence (probability theory)
Information retrieval
Mathematics
Roundness (object)
Data compression
Object (grammar)
Code
Information
Information security
Speech synthesis
Area
Algorithm
Touchscreen
Process (computing)
Spacetime
Temporal logic
Structural load
Sampling (statistics)
Staff (military)
Group theory
Demoscene
Network topology
Order (biology)
Right angle
Whiteboard
Digitizing
Row (database)
Point (geometry)
Frame problem
Readonly memory
Computergenerated imagery
Data recovery
Characteristic polynomial
Streaming media
Field (computer science)
Number
Performance appraisal
Causality
Energy level
Elearning
Subtraction
Units of measurement
Metropolitan area network
Standard deviation
Information
Direction (geometry)
Expert system
Trajectory
Shape (magazine)
Table (information)
Summation
Computer animation
Personal digital assistant
Information retrieval
Key (cryptography)
Object (grammar)
Game theory
Freezing
Matching (graph theory)
Separation axiom
1:40:10
Frame problem
Pattern recognition
Scaling (geometry)
Computergenerated imagery
Parameter (computer programming)
Angle
Transformation (genetics)
Translation (relic)
Element (mathematics)
Rotation
Sequence
Flow separation
Computer animation
Object (grammar)
Data compression
Information
Pairwise comparison
Data type
Design of experiments
1:45:54
Spacetime
Group action
Musical ensemble
Curvature
Ferry Corsten
Model theory
Multiplication sign
Insertion loss
Angle
Trajectory
Explosion
Object (grammar)
Flag
Physical law
Circle
Data conversion
Pairwise comparison
Information security
Social class
Area
Graphics tablet
Boss Corporation
Algorithm
Structural load
Bit
Staff (military)
Demoscene
Hand fan
Arithmetic mean
Network topology
Smart card
Data logger
Computer programming
Threedimensional space
Observational study
Presentation of a group
Virtual machine
Inclined plane
Mathematical analysis
Sound effect
Broadcasting (networking)
Goodness of fit
Term (mathematics)
Energy level
Right angle
Subtraction
Units of measurement
Shift operator
Information
Content (media)
Trajectory
Frame problem
Hypermedia
Computer animation
Information retrieval
Radiofrequency identification
Object (grammar)
Film editing
1:48:19
Frame problem
Information retrieval
Pattern recognition
Computer animation
Set (mathematics)
Presentation of a group
Group action
Abstraction
Demoscene
Sequence
1:53:08
Information retrieval
Pattern recognition
Computer animation
Hypermedia
View (database)
Scientific modelling
Group action
Mereology
Abstraction
Markov chain
1:54:46
Computer animation
Abstraction
Metadata
Formal Metadata
Title  Hidden Markov Model, Video Retrieval (09.06.11) 
Title of Series  Multimedia Databases 
Part Number  10 
Number of Parts  14 
Author 
Balke, WolfTilo

Contributors 
Homoceanu, Silviu

License 
CC Attribution  NonCommercial 3.0 Germany: You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal and noncommercial purpose as long as the work is attributed to the author in the manner specified by the author or licensor. 
DOI  10.5446/340 
Publisher  Technische Universität Braunschweig, Institut für Informationssysteme 
Release Date  2011 
Language  English 
Producer 
Technische Universität Braunschweig

Production Year  2011 
Production Place  Braunschweig 
Content Metadata
Subject Area  Information technology 
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 contentspecific 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, PrecisionRecall Analysis  Semantic content of imagecontent search  Image representation, lowlevel and highlevel features  Texture features, randomfield 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, MPEG7  Extraction of lowand highlevel features Integration of features and efficient similarity comparison  Indexing over inverted file index, indexing Gemini, R * trees 