Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
V i s u a l I n t e l l i g e n c e C o m p a n y © 2017. Deepixel Inc. all rights reserved.
모바일 VR 사용자 인터페이스를
위한 데이터 기반 기...
2
© 2017. Deepixel Inc. all rights reserved.
Contents
Data-driven Machine Learning
I n t e l l i g e n ce , M a c h i n e ...
© 2017. Deepixel Inc. all rights reserved.
Data-driven Machine Learning
Int elligence, M a chine Lea rning | Deep Lea rnin...
4
© 2017. Deepixel Inc. all rights reserved.
Intelligence
•
• Turing test: A test of a machine's ability to exhibit intell...
5
© 2017. Deepixel Inc. all rights reserved.
Intelligence
• How can we add intelligence to the machine?
Data
Algorithm/Rul...
6
© 2017. Deepixel Inc. all rights reserved.
Machine Learning
• Field of computer science that gives computers the ability...
7
© 2017. Deepixel Inc. all rights reserved.• https://machinelearningmastery.com/
A mind map of machine learning algorithm...
8
© 2017. Deepixel Inc. all rights reserved.
Deep learning
• Large neural network with a cascade of multiple layers of non...
9
© 2017. Deepixel Inc. all rights reserved.
Deep Learning X Big Data
Cat face, Google, 10M 200x200 captures from YouTube
...
10
© 2017. Deepixel Inc. all rights reserved.
Datasets for Machine Learning | Deep Learning
• https://en.wikipedia.org/wik...
11
© 2017. Deepixel Inc. all rights reserved.
THE MNIST(Modified National Institute of Standards and Technology) DATABASE
...
12
© 2017. Deepixel Inc. all rights reserved.• http://www.vision.caltech.edu/Image_Datasets/Caltech_10K_WebFaces/
Caltech ...
13
© 2017. Deepixel Inc. all rights reserved.
ImageNet
Labeled objects, SIFT features for object detection/recognition LSV...
14
© 2017. Deepixel Inc. all rights reserved.
Deep Learning w/ or w/o Big Data
• Deep learning is a powerful tool but, it ...
15
© 2017. Deepixel Inc. all rights reserved.
Data Collection for Deep Learning
• End-to-End Learning of Deep Visual Repre...
16
© 2017. Deepixel Inc. all rights reserved.
Data Collection for Deep Learning
• The SYNTHIA Dataset: A Large Collection ...
17
© 2017. Deepixel Inc. all rights reserved.
Deep Learning w/ or w/o Big Data (again)
• Deep learning is one of the best
...
© 2017. Deepixel Inc. all rights reserved.
Mobile VR User Interface
P I X I E , Reinvent ed 3D N at ura l Int erfa ce for ...
19
© 2017. Deepixel Inc. all rights reserved.
• https://vr.google.com/
• https://www.vive.com/
Mobile VR PC/Console VR
20
© 2017. Deepixel Inc. all rights reserved.
Mobile VR Interface
• https://uxdesign.cc/
21
© 2017. Deepixel Inc. all rights reserved.• http://edition.cnn.com/videos/world/2016/11/28/alibaba-vr-shopping-stevens-...
22
© 2017. Deepixel Inc. all rights reserved.
PIXIE Reinvented 3D Natural Interface for Mobile VR
23
© 2017. Deepixel Inc. all rights reserved.
Hand motion tracking in a VR/AR environment
• http://www.samsung.com/us/camp...
24
© 2017. Deepixel Inc. all rights reserved.
Our Approach
• Challenging Issues
• Difficult to collect the desired trainin...
25
© 2017. Deepixel Inc. all rights reserved.
DEEP(Data Extraction Equipment & Program)
26
© 2017. Deepixel Inc. all rights reserved.
DEEP dataset
Hand and fingertip action videos and images for PIXIE
50+ video...
27
© 2017. Deepixel Inc. all rights reserved.
Algorithm
• https://en.wikipedia.org/wiki/Local_binary_patterns
• https://en...
28
© 2017. Deepixel Inc. all rights reserved.
C
https://www.youtube.com/watch?v=aCP6__kv1zM
29
© 2017. Deepixel Inc. all rights reserved.
Concluding Remarks
• Design suitable methods and choose the right machine le...
30
© 2017. Deepixel Inc. all rights reserved.w w w . d e e p i x e l . x y z
Thank You
Upcoming SlideShare
Loading in …5
×

of

모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 1 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 2 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 3 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 4 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 5 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 6 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 7 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 8 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 9 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 10 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 11 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 12 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 13 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 14 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 15 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 16 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 17 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 18 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 19 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 20 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 21 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 22 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 23 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 24 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 25 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 26 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 27 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 28 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 29 모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표 Slide 30
Upcoming SlideShare
What to Upload to SlideShare
Next
Download to read offline and view in fullscreen.

0 Likes

Share

Download to read offline

모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표

Download to read offline

2017 데이터 그랜드 컨퍼런스 발표 자료

  • Be the first to like this

모바일VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 - 딥픽셀 이제훈 대표

  1. 1. V i s u a l I n t e l l i g e n c e C o m p a n y © 2017. Deepixel Inc. all rights reserved. 모바일 VR 사용자 인터페이스를 위한 데이터 기반 기계 학습 이제훈 2017 데이터 진흥주간 | 2017 데이터 그랜드 컨퍼런스 | 지능형 데이터 for AI | 2017년 11월 7일
  2. 2. 2 © 2017. Deepixel Inc. all rights reserved. Contents Data-driven Machine Learning I n t e l l i g e n ce , M a c h i n e Le a r n i n g | D e e p Le a r n i n g , B i g D at a Mobile VR User Interface P I X I E , Re i n v e n t e d 3 D N at u r a l I n t e r f a ce f o r M o b i le V R Concluding Remarks P18 P3 P29
  3. 3. © 2017. Deepixel Inc. all rights reserved. Data-driven Machine Learning Int elligence, M a chine Lea rning | Deep Lea rning, Big Dat a
  4. 4. 4 © 2017. Deepixel Inc. all rights reserved. Intelligence • • Turing test: A test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human [Alan Turing, MIND, 1950] • https://en.wikipedia.org/wiki/Turing_test
  5. 5. 5 © 2017. Deepixel Inc. all rights reserved. Intelligence • How can we add intelligence to the machine? Data Algorithm/Rule Output Machine Machine Algorithm/Rule Data Output Machine LearningTraditional Programming C
  6. 6. 6 © 2017. Deepixel Inc. all rights reserved. Machine Learning • Field of computer science that gives computers the ability to learn without being explicitly programmed (Arthur Samuel, 1959) • Data-driven approach: Generalization and pattern discovery from training data • Data is crucial to machine learning. • Traditional programming • Impossible to make a perfect rule Too many rules Too many factors influencing the rules Obscure rules • https://en.wikipedia.org/wiki/Machine_learning
  7. 7. 7 © 2017. Deepixel Inc. all rights reserved.• https://machinelearningmastery.com/ A mind map of machine learning algorithms (by type)
  8. 8. 8 © 2017. Deepixel Inc. all rights reserved. Deep learning • Large neural network with a cascade of multiple layers of nonlinear processing units • Improvement of • Algorithms: Unsupervised pre-training [G. Hinton, TCS, 2007], Dropout [G. Dhal, ICASSP, 2013], and so on. • H/W : GPU • Big data: Lots of datasets from the Internet, SNS, and so on.
  9. 9. 9 © 2017. Deepixel Inc. all rights reserved. Deep Learning X Big Data Cat face, Google, 10M 200x200 captures from YouTube Large scale unsupervised learning [Quoc V. Le et al., ICML, 2012] AlphaGo, Google DeepMind, ~30M moves, large numbers of games against other instances of itself Deep learning w/ supervised learning and reinforcement learning [D. Silver et al., Nature, 2016] • https://googleblog.blogspot.kr/2012/06/using-large-scale-brain-simulations-for.html • http://www.nature.com/nature/journal/v529/n7587/index.html
  10. 10. 10 © 2017. Deepixel Inc. all rights reserved. Datasets for Machine Learning | Deep Learning • https://en.wikipedia.org/wiki/List_of_datasets_for_machine_learning_research • FERET (facial recognition technology) • CMU Pose, Illumination, and Expression (PIE) • YouTube Faces DB • Yale Face Database • TV Human Interaction Dataset • Berkeley Multimodal Human Action Database (MHAD) • Microsoft Common Objects in Context (COCO) • ImageNet • MNIST database • Stanford Dogs Dataset • Microsoft Sequential Image Narrative Dataset (SIND) • Berkeley 3-D Object Dataset • Etc.
  11. 11. 11 © 2017. Deepixel Inc. all rights reserved. THE MNIST(Modified National Institute of Standards and Technology) DATABASE Handwritten digits for handwriting and character recognition 60,000 images • http://yann.lecun.com/exdb/mnist/
  12. 12. 12 © 2017. Deepixel Inc. all rights reserved.• http://www.vision.caltech.edu/Image_Datasets/Caltech_10K_WebFaces/ Caltech Web Faces Frontal face images for facial recognition 10,000 images
  13. 13. 13 © 2017. Deepixel Inc. all rights reserved. ImageNet Labeled objects, SIFT features for object detection/recognition LSVRC (Large Scale Visual Recognition Challenge) 14,197,122+ images (Y2010) • http://image-net.org/index
  14. 14. 14 © 2017. Deepixel Inc. all rights reserved. Deep Learning w/ or w/o Big Data • Deep learning is a powerful tool but, it requires lots of annotated data. • Annotated data may be • Expensive (human labor) • Very difficult to collect • Not publicly available • How to effectively collect large scale data with little human labor? • Data augmentation • Synthetic data • Data cleaning • Etc.
  15. 15. 15 © 2017. Deepixel Inc. all rights reserved. Data Collection for Deep Learning • End-to-End Learning of Deep Visual Representations for Image Retrieval [Gordo et al., IJCV, 2017] To automatically clean datasets by pruning edges of matched images with the pairwise matching graph from public datasets with noisy. • Frankenstein: Learning Deep Face Representations using Small Data [Hu et al., TIP, 2017] To generate very large training datasets of synthetic images by composition of real face images in a given dataset, which swaps the facial components of different face images to generate a new face.
  16. 16. 16 © 2017. Deepixel Inc. all rights reserved. Data Collection for Deep Learning • The SYNTHIA Dataset: A Large Collection of Synthetic Images for Semantic Segmentation of Urban Scenes. [Ros et al., CVPR, 2016] To render a virtual city created with the Unity development platform and automatically generate realistic synthetic images with pixel-level annotations. • Data-Augmentation for Reducing Dataset Bias in Person Re-identification [McLaughlin et al., AVSS, 2015] Random modification including background simulation of the raw data to increase dataset diversity. C
  17. 17. 17 © 2017. Deepixel Inc. all rights reserved. Deep Learning w/ or w/o Big Data (again) • Deep learning is one of the best methods in machine learning. • Big data is crucial to deep learning. • Data scarcity problem • Alternative methods for collecting data
  18. 18. © 2017. Deepixel Inc. all rights reserved. Mobile VR User Interface P I X I E , Reinvent ed 3D N at ura l Int erfa ce for M obile VR
  19. 19. 19 © 2017. Deepixel Inc. all rights reserved. • https://vr.google.com/ • https://www.vive.com/ Mobile VR PC/Console VR
  20. 20. 20 © 2017. Deepixel Inc. all rights reserved. Mobile VR Interface • https://uxdesign.cc/
  21. 21. 21 © 2017. Deepixel Inc. all rights reserved.• http://edition.cnn.com/videos/world/2016/11/28/alibaba-vr-shopping-stevens-pkg.cnn
  22. 22. 22 © 2017. Deepixel Inc. all rights reserved. PIXIE Reinvented 3D Natural Interface for Mobile VR
  23. 23. 23 © 2017. Deepixel Inc. all rights reserved. Hand motion tracking in a VR/AR environment • http://www.samsung.com/us/campaign_smarttv/ • https://www.microsoft.com/en-us/hololens/ Smart TV AR Headset
  24. 24. 24 © 2017. Deepixel Inc. all rights reserved. Our Approach • Challenging Issues • Difficult to collect the desired training dataset: the back of the hand designed shape of the hand for UX/UI dataset with various colors and poses of the hand • 3D pose estimation with no additional device other than a smartphone with a single camera • Real-time operation ( 60ms) in a mobile environment • Our Approaches • Development of DEEP(Data Extraction Equipment and Program) to collect data • -big- • Algorithm optimization to reduce computational complexity
  25. 25. 25 © 2017. Deepixel Inc. all rights reserved. DEEP(Data Extraction Equipment & Program)
  26. 26. 26 © 2017. Deepixel Inc. all rights reserved. DEEP dataset Hand and fingertip action videos and images for PIXIE 50+ video files 1h+ running time 20,000+ images
  27. 27. 27 © 2017. Deepixel Inc. all rights reserved. Algorithm • https://en.wikipedia.org/wiki/Local_binary_patterns • https://en.wikipedia.org/wiki/Support_vector_machine
  28. 28. 28 © 2017. Deepixel Inc. all rights reserved. C https://www.youtube.com/watch?v=aCP6__kv1zM
  29. 29. 29 © 2017. Deepixel Inc. all rights reserved. Concluding Remarks • Design suitable methods and choose the right machine learning algorithm for your problem according to the size of datasets you have, performance of available hardware, and challenging issues of the given problem. • Collect your dataset with a clever way , if you choose a machine learning algorithm as your core solution.
  30. 30. 30 © 2017. Deepixel Inc. all rights reserved.w w w . d e e p i x e l . x y z Thank You

2017 데이터 그랜드 컨퍼런스 발표 자료

Views

Total views

279

On Slideshare

0

From embeds

0

Number of embeds

0

Actions

Downloads

1

Shares

0

Comments

0

Likes

0

×