How to apply deep learning for 3D object
Formal Metadata
Title 
How to apply deep learning for 3D object

Title of Series  
Author 

License 
CC Attribution  NonCommercial  ShareAlike 3.0 Unported:
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 and the work or content is shared also in adapted form only under the conditions of this license. 
Identifiers 

Publisher 

Release Date 
2017

Language 
English

Content Metadata
Subject Area  
Abstract 
How to apply deep learning for 3D object [EuroPython 2017  Talk  20170713  Arengo] [Rimini, Italy] I talk about the ""How to achieve the 3D object recognition accuracy 80%(40 category) for 3month " Deep Learning is the good technique for image recognition and speech recognition. And it apply the other field. Many people try to apply the Deep Learning, but it is difficult to make a result. In my situation, I have enough knowledge about the 3D object and label data. I'll talk about the how to achieve the 80 % (40 category) In My approach 1: Getting the Information 1.1: How to choose the information 1.2: How to choose framework 2: Getting the Data 2.1: Public data 2.2: How to make the own data 3: Try small 3.1: Trying the small data set 3.2: Trying the train and predict 4: Deciding the direction focus 4.1: Choose what you can control 5: Prioritizing with high certainty 5.1: Preprocess 5.2: Improve the train speed 6: Increasing the challenge times 6.1: Using the GPU 6.2: CPU optimization 6.3: multi process 6.4: resource 7: Parameter Tuning 7.1: Improve Model Versatility or Improve Data Versatility 7.2: Model Tuning 7.2.1: RandomDropOut 7.2.2: LeakyRelu 7.3: Data Argumantion 8: Product 8.1: Minimum function 8.2: Using Docker I hope to people who want to apply Deep Learning for the 3D mode

00:00
Intel
Presentation of a group
Object (grammar)
Plotter
Software
Personal digital assistant
Strategy game
Mereology
Object (grammar)
Mereology
Area
Product (business)
00:30
Focus (optics)
Computer virus
Table (information)
Software developer
State of matter
Weight
Image processing
1 (number)
Content (media)
Focus (optics)
Wave packet
Twitter
Data model
Process (computing)
CNN
Function (mathematics)
Personal digital assistant
Strategy game
Right angle
Information
Traffic reporting
Resultant
01:09
Covering space
Observational study
1 (number)
Determinant
Mereology
01:33
Pattern recognition
Image processing
Time series
Mathematical analysis
Formal language
Product (business)
Process (computing)
Personal digital assistant
Series (mathematics)
Personal digital assistant
Natural number
Process (computing)
Speech synthesis
02:25
Googol
Googol
Right angle
02:54
Group action
Content (media)
Code
System call
Twitter
Computer programming
Structured programming
Goodness of fit
Googol
Googol
Right angle
Information
Data structure
Identical particles
Logic gate
Metropolitan area network
04:03
Boss Corporation
Multiplication sign
Calculation
Visual system
Set (mathematics)
Grass (card game)
Product (business)
Wave packet
Product (business)
Tensor
Befehlsprozessor
Whiteboard
Befehlsprozessor
Gastropod shell
Speech synthesis
Formal verification
Condition number
Graphics processing unit
05:12
Data model
Structured programming
Preprocessor
Parameter (computer programming)
Mathematical optimization
Data Augmentation
Mathematical optimization
Product (business)
05:37
Addition
Random number
Axiom of choice
Linear regression
Support vector machine
Weight
Multiplication sign
Plotter
Logistic distribution
Focus (optics)
Product (business)
Product (business)
Forest
Preprocessor
Object (grammar)
Personal digital assistant
Data Augmentation
06:26
Source code
Observational study
State of matter
Software developer
Multiplication sign
Code
Sound effect
Demoscene
Product (business)
Data model
Type theory
Integrated development environment
Readonly memory
Core dump
Authorization
output
Cuboid
Integrated development environment
Force
Reduction of order
Graphics processing unit
07:52
Data model
Convolution
CNN
Personal digital assistant
Object (grammar)
Computergenerated imagery
1 (number)
output
Species
output
Window
08:49
Data model
Convolution
CNN
Object (grammar)
Multiplication sign
Sheaf (mathematics)
Figurate number
output
09:21
Convolution
State of matter
Multiplication sign
Content (media)
Set (mathematics)
Sound effect
Maxima and minima
System call
Power (physics)
Data model
Arithmetic mean
Causality
CNN
Personal digital assistant
Object (grammar)
Right angle
Analytic continuation
Spacetime
10:32
Convolution
Data model
Group action
File format
Window
11:00
Asynchronous Transfer Mode
Distribution (mathematics)
Arm
Graph (mathematics)
Distribution (mathematics)
Image resolution
Chaos (cosmogony)
Product (business)
Data model
Population density
Personal digital assistant
Function (mathematics)
Normal (geometry)
Right angle
Quicksort
11:46
Functional (mathematics)
Metric system
Interface (computing)
Set (mathematics)
Insertion loss
Shape (magazine)
Wave packet
Product (business)
Product (business)
Sequence
Data model
Kernel (computing)
Website
output
Metric system
12:24
Shift operator
Variety (linguistics)
Weight
Multiplication sign
Software developer
Shared memory
Bit
Augmented reality
Limit (category theory)
Computer font
Product (business)
Product (business)
Data model
Category of being
Mathematics
Vector space
Personal digital assistant
Speech synthesis
Social class
Right angle
Abelian category
Data Augmentation
13:47
Rotation
Personal digital assistant
Computergenerated imagery
Bit
Special unitary group
Data Augmentation
Cartesian coordinate system
Rotation
Number
Physical system
14:35
Befehlsprozessor
Readonly memory
Befehlsprozessor
Calculation
Software testing
Mathematical optimization
Multiplication
Thread (computing)
Graphics processing unit
15:08
Computer configuration
Computer configuration
Befehlsprozessor
Calculation
Mathematical optimization
Resultant
15:28
Addition
Arithmetic mean
Line (geometry)
Weight
Weight
Shift operator
Data Augmentation
Resultant
Physical system
Wave packet
16:24
Pattern recognition
Bit rate
Strategy game
Strategy game
Right angle
Limit (category theory)
Computer programming
17:04
Point (geometry)
Permanent
Demo (music)
Demo (music)
Maxima and minima
Website
Speech synthesis
Object (grammar)
Gamma function
Product (business)
Installable File System
19:20
Electronic data interchange
Demo (music)
Graphic design
Shape (magazine)
20:53
Group action
Object (grammar)
Demo (music)
21:15
Rotation
Computer virus
Machine learning
Presentation of a group
Object (grammar)
Physical law
Matrix (mathematics)
Representation (politics)
Right angle
Metric system
Disk readandwrite head
Social class
00:05
so think you hope have coming to my presentation and I talk about how to right people running to see the objects
00:14
so this physical consists of a 2 part of + plot you'd get knowledge about how to approach to making the the running crawled out the 2nd plot you get knowledge about people earning up right to 3 the opposite
00:32
so the 1st and consist of about these things the start to the ones that I grow well find the right nestled on the keeper really challenging focus the 2nd 1 states that the branding upright is the opposite of adult cobalt of the books net on the holding bruited books that on the results this is my favorite
00:58
that so my name is maasai oversee I walked up public income that I'm I'm is process in the report if you think might be the mighty ducks out what is the call think
01:12
cover pink provided to the on ondemand the morning studies sometimes receives the threeday determin Hawkstone only so by using 3 printers and other so that the reason why ital cobalt deep rounding up right to see the overseer so that
01:31
was part of the entity ones that I for
01:35
1 we do that these deeper learning to solve
01:40
a problem is you no thanks so the
01:48
work to the best gift you keep running so the most cases imageprocessing house political emission of high home than the other missile as some some cases it is possible it of among processing and time series analysis what is the worst case do use deep running so not enough data to calculate their time although on the need to 100 football and aggressive because ingest became a synonym possibility that it is offered to was 1 100 % accuracy so
02:27
the next is the 1 the right to missile How can find the best
02:34
way to so problem using the binding in the mall scares use that leads to such not possible because so that you use the google you have to all about the Duke you out so the art past you don't know about that but you
02:54
a man quality the using Google Scholar so it is to get to the hallway things go Nestle's who Q of the team but they're about this known about these programs so you already know the BG busted oversight is so it is possible to get the cold and data so you might get some pretty staying busty provided the data and called and the DCT gates cities that it provided that the type of on the right is the paper of so you really get to the cold and paper on the the data so it is possible to get so distinguishes and there's so the book you you get the structure of knowledge and at it is provided the paper cites so the demand of a dismissal to and and this so you already know the group you also the you you you the harbor on google it is did was that they get a good call and good knowledge so the
04:00
next content they keep the challenging so you get the
04:05
jails that are a lot the training data so let's start using the through this set not possible both so that a lot of Atlanta dating several time to that alright so you have to check that was
04:20
seen of passing so bell small dataset and take Moser what worked 3 times cost that would be a need to better apply 2 innings dataset in most models can be 2 armies datasets such of them you have to take it to what the on there the planning out how to value not all shells missile Tampa accuracy so you have to check that for any accuracy on the body's Juncker C is the boss accuracy of melting Brouard so you have to stop it to undertake the result by jumbled such of that is of all on the health time times while the varying in the conditions speech is in GPU optimize CPU yeah whole grass
05:16
the banning of real Tomaso the improve accuracy so the model of how the how easy to structure adjusting the hyperparameter on the preprocessed data the documentation either using graphical data on the optimizer is the 1 song
05:38
so depending on the newest addition so the inner competition associated on the enough data for the complex so we have refills but enough data to find a good free time all and focus on the proposes such as delimitation so not you have competition resources or data consider other weights so liberal programs such as the model was the node is no is the creation of Thomas Mann none of the left and the deepening prowhite these and best stories so the end of the past
06:21
plot of a talk about next spot the Pirani applied to 3 goes on the I
06:28
don't what books there are a lot of the
06:33
primal oath how to choose among
06:36
I consider scene the souls of the home of speed solely so the that competition authorities and human so the all of the the accuracy and speed speed of development I choose the books that right to the books that's a box they don't want is resource on speed so competition itself is good because they might environment and to walks namely sadly together by on the GPU types is there on the home of the that there's the also they deceive authentic guessing In the paper and speed is open source and in the core so I talk about books meant to prophecies at half time so them out the city did that subject to subject to subject the book fair so the study taking the possibility to choose 8 under the states that because the US 3 it's data so the we have to reduce the size and the input to the Congress also the so the Commerce and see if effective of there expand about the the kisses
07:56
for because the the the case is difficult so the BPA the input team s and the but there's a kind of window so you really want to some of its species around the In the detainees bodies along and the ones from the 5 and trust and all in the team is bodies along almost 4 1 and trust repeated outstanding on the it is get to the body Congress refused so 3 on the repeated so when species greenspaces the and the kids get so Mr. future body and there's
08:40
also talk about the city cases having provided the and that the manners and the city the and the economy and do
08:52
that in its on the loading is not on the provided to commission future on the news saying now and the related to a commissioner future the the duck sounding so the get that fast Commission
09:12
future so in in Figure it is set up time section so if you
09:23
think the 15th entering into the call so the fees that can it is only 1 minor so in this case fact it is so this in that case in that state is the subject is that it started in the left while 1 meaning about the current time under the fitting the corner on the societies that have the continuity and they go home the meaning about to the beach content so that care support of all that is the role of ASEAN also the principle unsound or a divine being safe so that the reason why the setting that power the current so the next step
10:07
is the max through some marks effective cause the the dictate immense and then so keep to get to that Commission future right here on the so Sikhism exploded In that infinite space the possible in the 6th so the get 6 ongoing spaces it's among a baby to same optional it the next
10:35
thing action for the sake that it is the same action so the provided to the following window on the promise of future the deduction it is amongst through only I'm going
10:53
to use the can is only so 1 liner soliciting size and Deltona
11:03
so the entertainment is so there are fewer dying undertakes arms over require heat and side the any decided because of the in the crust fly cases normal graph on the right the filter months function that's so smart postulate a MoPPS out food to provide the distribution on the the it is easy to define so the chaos using is
11:33
only these gains things so the they have about of through a connected on images size immediate enable crust size on the sort amongst year that the goal is
11:51
to define the more there but for all the that divide the model on the homeless and for the owners of Treaty on the monks through and the fury connected imitate sites interface and so a bright softmax function so it
12:08
also define what is a loss function on the metrics and then training the data so the setting of the books that are not being data on those sitting the cross of so
12:26
the idea talk about the next step is same baby in taking a aggressive I
12:33
thinking baby accuracy has to have properties of hostile qualities of more than 1 2nd the present data so model quality of by the way to improve accuracy development days but the more that it's our old were resources and it is optional speech more better and date approach the what wanted to have active change of shift and of the this approach are R&D the so you may get more the properties of upright to run them wildly given the on the data properties data documentation falsities it up on graphs wait for all last but they worry data so say he did all and assigns a specific case so I will talk a bit about what I think we did the limitation of the of font as a wealth of the missile that vector of shares on the done nothing grazing collocation time unlike adding right to a model so the Government this of this thing so that attending data quotations shift here so
13:49
the creator applied is cold so that provided the books that a bit on the each books in data augment osmotic external that Britain getting the attending data on the weekends examined by wholemount
14:09
so talk about the number of cases In the rotation the sun so that ending that days of rotation system operates attending the data song and understand other x 10 years that they've got some so you might cancels the
14:39
findings of the momentous it doubles the during the so impairment techniques
14:49
speed the banning of the oral to
14:53
wasting bearing carcasses speed is sent to the use of CPU or CPU optimize its own lots well to prepare the test set so optimizes some very effective for the deducted ornamentation so we talk about CPU so use that is a full
15:12
also thinking the be adoption is posted up by the a city optimize have you have to take some of its option is available the so the
15:28
result this it is 0 so the by this
15:32
accuracy for the yeah I mean baseline the meaning about the date of an additional system that's on the system systemwide on the yeah I mean you know about the system x y on a bright across weight and the green meaning that about findings the training data are fooled by the the documentation on the bright across weight they stay Visual about 2 that
16:06
gives the results so this is the result of the additional literacy so the baselines in that 79 often have also the findings this x and y data on the bright across weights but see this 85 often so the
16:26
conclusion so the market
16:31
fusspots my strategy find the right mescal on the lower right when the right people like Wainwright rollout find right nestled on the Chinese and focus so the so the right of program is saying the opposite to the Galician and alright to like cheese books that on the low rate chinese ink data limitations on the modeled on the focusing Britain and by this accuracy others
17:08
show the demo what then on these sites show of all of the Sumerian of 3 D objects have also these speech to those of an yes at this point this the so I think some of the permanent repeal year again shows a beam it is quite well to
19:01
the brain under set yeah you OK for but what the so early
19:35
so they're trying these posts refined signal shape on that and the next the quantities that passed for the rest quantity that erect have altered it down the TAT possible fine a such a so the end of his apprentice
21:08
armed also has a group think I think in the primal the 2 and those who
21:17
in the theme the logging in Japan so you have to to society and all it will take you seem as a now so all think you're risking my presentation if the is the head of the virus so so questions 1 has and so on hello representations and you hear any other metrics of just stressed because accuracy especially if you have imbalanced classes be produced the right the law and lost or a precision recall what about so you can assume you know about the other that countries symmetric so such that their rotation and so on the this the the come from men of the matrix I so I feel that 0 I only then the packet about the accuracy so not come from no not calculated of that matrix and any other questions but it will consist of