Viewdle is a Silicon Valley start-up specializing in face recognition technology for media indexing, search, and navigation using advanced video analysis techniques. The company aims to improve user experiences in media management and augmented reality by offering fast, accurate, and scalable solutions. Viewdle is looking to hire talented individuals in image processing and machine learning to further enhance their innovative technologies.

1. Face recognition
for augmented reality and
media management

2. Viewdle
Face recognition based
media indexing, search,
navigation and editing

3. Viewdle: center of excellence for videos
A silicon valley start-up in Ukraine and in California
VC backed company Specialists in video analysis
World renown scientist in 4 patents filed in the US
Structural Recognition
Professor Schlesinger
Culture of innovation:
Environment that fosters
40 Engineers, 8 PhDs. All with creativity, and dedication
Advance degrees in
Mathematics and Computer
Science Head-quartered in Palo-Alto

5. Media indexing
We find
people in
photos & videos

6. Performance
 Accurate
 Fast
 Scalable
 Easy to use

10. Viewdle Video Indexing
We find
people in
video

11. Why we search people in video?
 People names are one of
most typical words used in
search queries

12. Performance
 Accurate
 Fast
 Scalable
 Easy to use

13. Video indexing workflow
Video 1 indexing &
Manual Other modalities
faces agglomerative
face tagging
clustering
Person DB Fusion &
population Search index
creation
Video 2 indexing & Automatic face
faces agglomerative suggestion/tagging
clustering (face recognition)

14. Search workflow
Search within index Play video Provide
of tagged or recognized from moment of navigation by
videos person appearance person appearances

15. Technology applications
 Video& photo archive search
 Personal media search
 Improved media sharing
experience for desktop and
mobile
 Augmented reality
 Video editing

16. Technology applications
continued
 Search engine optimization
 Content filtering
 Security

17. Face based search advantages
 More relevant results than
from text
 Pointing to exact position in
video

18. Not relevant OCR

19. Problems in video
 Low video quality (Youtube 420x300)
 Low quality faces even in high
quality videos
 Complex lighting, facial expression,
pose, occlusion (microphone or cup)‫‏‬
 Large # of persons to recognize
 Similar looking people

20. Digital TV not ideal

21. Faces clustering challenge

22. How we solve problems?
 Cascaded detectors
and recognizers
 Large set of manually created
ground truth data(> 100k
images)
 Automated detectors and
recognizers thresholds
optimization

23. Video Processing Performance
 Accuracy: Recall 65+%, Precision
95+%
 Realtime processing

24. Mobile targets
 Realtime 2D face tracking
 Online and offline face recognition
 Realtime 3D face tracking (future)

25. Teaser video

26. Mobile Challenges
 Realtime video stream processing
 Less powerfull processors (ARM
1GHZ)
 Less capable OS (Android)

27. Desktop vs Mobile
 Every target platform requires
specific tuning and optimization for
best user experience
 Single algorithmic C++ code base

28. Future mobile hardware
 Qualcomm 4 core 2.5GHz
Snapdragon ARM processor
 Texas Instruments OMAP 5 platform
 nVidia 4 core ARM with GPU
 AMD 2 core Bobcat with GPU
 Intel Atom for phones

29. Reuters Demo
 http://reuters.viewdle.com

30. Search page

31. Exact person selection

32. Video selection

33. Opening video at position

34. Video navigation by people

36. Photo vs Video challenges
 Face detect is slower (images are bigger
dimensions, good faces can be smaller)
 Rotated and profile faces are harder to pick up
(in videos they frequently can be tracked from
frontals)
 More reliable face feature detection required
 Less # of faces require better face similarity
function
 Face clustering heuristics are more important

37. Photo improvements done
 Added profiles
 Improved speed
 Multi-cascaded scheme for different rotations
 Improved general recall/precision
 Added face grouping heuristics
 Bring to production Photo-Video SDK
 Mobile specific tech tuning

38. Photo SDK demo

39. Mobile Demo

40. Things to improve
 Face detection recall/precision (improving
training set)
 Feature detection improvement
 Profiles recognition
 Clustering heuristics tuning
 Speed improvement

41. Conclusion
 Viewdle provides quality
product & technology for media
indexing, search and navigation
 Viewdle creates new user
experiences never existed
before

42. Looking for talents
 Viewdle invites young scientists
and engineers interested in image
processing, machine learning,
object tracking to work at Viewdle
 Send application to
jobs@viewdle.com or
 ysm@viewdle.com

43. Thank you for your attention!
Yuriy Musatenko, Ph.D., CTO.
Ivan Kovtun, Ph.D., Head of research.
Viewdle, Inc.