1. Mixed Scale Motion Recovery James Davis Ph.D. Oral Presentation Advisor – Pat Hanrahan Aug 2001

2. 2 High level goal Recover motion Recover motion Large working volume Extreme detail

3. 3 Acquire real motion as computer model Acquire real motion as computer model Fixed resolution vs. range ratio Fixed resolution vs. range ratio Current technology

4. 4 Applications of motion recovery Animation Animation Athletic analysis Athletic analysis Biomechanics Biomechanics

5. 5 Mixed scale domains Detailed motion within a larger volume Detailed motion within a larger volume

6. 6 Problem domain characterization Multiple scales of motion Multiple scales of motion At individual scales At individual scales Working volume is local Working volume is moving

7. 7 Hierarchical paradigm Explicitly expresses motion hierarchy Explicitly expresses motion hierarchy Motion recovery drives sub-region selection Sub-region selection defines next scale Multiple designs possible within framework Multiple designs possible within framework Motion recovery Sub-region selection Motion recovery Sub-region selection Large scaleMedium scaleSmall scale

8. 8 My design Multi-camera large scale recovery Multi-camera large scale recovery Covers large volume Robust to occlusion Pan-tilt multi-camera small scale recovery Pan-tilt multi-camera small scale recovery Automated camera control High resolution imaging

9. 9 Demonstration of my system Body moves on room size scale Body moves on room size scale Face deforms on much smaller scale Face deforms on much smaller scale Simultaneous capture Simultaneous capture

10. 10 Desirable system properties High resolution/range ratio High resolution/range ratio Scalable Scalable Occlusion robustness Occlusion robustness Runtime automation Runtime automation

11. 11 Related Work Traditional motion recovery Traditional motion recovery [Vicon] [MotionAnalysis] [Guenter98] Simple pan/tilt systems Simple pan/tilt systems [Sony EVI-D30 ] [Fry00] 2D guided pan/tilt systems 2D guided pan/tilt systems [Darrell96] [Greiffenhagen00] Human controlled cameras Human controlled cameras [Kanade00] Runtime automation High resolution/range ratio ScalableOcclusion robustnessRecovers motion

12. 12 Contributions Framework for mixed scale motion recovery Framework for mixed scale motion recovery Hierarchical paradigm Data driven analysis Model based solutions Specific system design Specific system design High resolution/range ratio Scalable Robust to occlusion Automated Application to simultaneous face-body capture Application to simultaneous face-body capture

13. 13 Talk outline Introduction Introduction Framework Framework Hierarchical paradigm Data driven analysis Model based solutions System implementation System implementation Large scale recovery Sub-region selection Small scale recovery End-to-end video End-to-end video Summary and future work Summary and future work

14. 14 Relation of system to hierarchy Large scale motion recovery Small scale motion recovery Sub-region selection

15. 15 System overview Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Large scale motion recovery Small scale motion recovery Sub-region selection

16. 16 Interface is the challenge Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Large/small scale similar Large/small scale similar Interface requirements differ Interface requirements differ Interface critical in end-to-end system Interface critical in end-to-end system Often ignored in individual components Often ignored in individual components Interfaces

17. 17 Data flow Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points System viewed as data flow System viewed as data flow Clean interface (data) desirable Clean interface (data) desirable

18. 18 System challenges Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Occlusion Noise Unknown correspondence Incorrect camera model Latency Occlusion Unknown camera motion

19. 19 Model improves data Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Occlusion Noise Unknown correspondence Incorrect camera model Latency Occlusion Unknown camera motion Model

20. 20 Model improves data Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Occlusion Noise Unknown correspondence Incorrect camera model Latency Occlusion Unknown camera motion Kalman filter P/T camera model Prediction Face model

21. 21 Talk outline Introduction Introduction Framework Framework Hierarchical paradigm Data driven analysis Model based solutions System implementation System implementation Large scale recovery Sub-region selection Small scale recovery End-to-end video End-to-end video Summary and future work Summary and future work

22. 22 Large scale system Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points

23. 23 Large scale physical arrangement 18 NTSC cameras 18 NTSC cameras 18 digitizing Indys 18 digitizing Indys

24. 24 Large scale features

25. 25 Large scale pose recovery Consider rays through observations Consider rays through observations Rays cross at 3D feature points Rays cross at 3D feature points

26. 26 Calibrating wide area cameras Jointly calibrated multiple cameras Jointly calibrated multiple cameras Iteratively estimate Iteratively estimate Camera calibration Target path [Chen, Davis 00]

27. 27 Unknown correspondence Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Unknown temporal correspondence Kalman filter

28. 28 Unknown temporal correspondence Multiple 3D features recovered Multiple 3D features recovered Which feature is the head? Which feature is the head? Each frame is independently derived Each frame is independently derived

29. 29 Dynamic motion model Not single frame triangulation Not single frame triangulation Dynamic motion model Dynamic motion model Model continuous motion Update on each observation Estimate position/velocity Extended Kalman filter [Kalman 60] [Broida86] [Welch97]

30. 30 Benefits of motion model Feature IDs maintained Feature IDs maintained Robust to short occlusion Robust to short occlusion Synchronized cameras unnecessary Synchronized cameras unnecessary

31. 31 Sub-region selection Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points

32. 32 Simplistic camera model Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Incorrect camera model P/T camera model

33. 33 Simplistic camera model Pan/tilt axes not aligned with optical center Pan/tilt axes not aligned with optical center IxIyIxIy = C R y R x xyzxyz

34. 34 New camera model Arbitrary pan/tilt axes Arbitrary pan/tilt axes Jointly calibrate axes and camera Jointly calibrate axes and camera Observe known points from several pan/tilt settings Fit data with minimum error = C T pan R pan T pan T tilt R tilt T tilt IxIyIxIy xyzxyz [Shih 97]

35. 35 Latency Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Latency Prediction

36. 36 Camera motor latency Empirically found 300ms latency Empirically found 300ms latency High velocity targets leave frame High velocity targets leave frame Prevents accurate sub-region selection Prevents accurate sub-region selection

37. 37 Target motion prediction Predict future target motion Predict future target motion Point camera at predicted target location Point camera at predicted target location Use previous motion model Use previous motion model High velocity objects successfully tracked High velocity objects successfully tracked P’ = P i +  t V i

38. 38 Small scale system Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points

39. 39 Small scale physical arrangement 4 Pan-tilt cameras point at sub-region 4 Pan-tilt cameras point at sub-region 4 SGI O2s digitize video 4 SGI O2s digitize video

40. 40 Small scale features Painted face features Painted face features Image gradient feature tracking Image gradient feature tracking [Lucas,Kanade 81] [Tomasi, Kanade 91]

41. 41 Small scale pose recovery

42. 42 Problems with face recovery Feature Tracking 3D Pose Recovery Video streams 2D points Pan/tilt parameters Pan/tilt controller LEDs 3D points Feature Tracking 3D Pose Recovery Video streams 2D points 3D points Occlusion Unknown camera motion Face model

43. 43 Problems with face recovery Self occlusion Self occlusion Many points not visible Camera motion not known precisely Camera motion not known precisely Difficult to merge more than two views Recovered 3D geometryView from one camera

44. 44 Face model Face model defines the set of valid faces Face model defines the set of valid faces Linear combination of basis faces Capture basis set under ideal conditions Basis transformation F =  w i B i F =  w i B i = w1w1 w2w2 w3w3 ++ [Turk, Pentland 91] [Blanz,Vetter 99] [Guenter et.al. 98]

45. 45 Model evaluation Mean error < 1.5 mm Remove features Fit to model Reconstruct features Observe everything Evaluate error

46. 46 Reconstructed face Fit partial data to the model Fit partial data to the model Use model to reconstruct complete geometry Use model to reconstruct complete geometry Recovered geometry from video Reconstructed geometry from model

47. 47 End to end video

48. 48 Talk outline Introduction Introduction Framework Framework Hierarchical paradigm Data driven analysis Model based solutions System implementation System implementation Large scale recovery Sub-region selection Small scale recovery End-to-end video End-to-end video Summary and future work Summary and future work

49. 49 Summary of contributions Framework for mixed scale motion recovery Framework for mixed scale motion recovery Hierarchical paradigm Data driven analysis Model based solutions Specific system design Specific system design High resolution/range ratio Scalable Robust to occlusion Automated Application to simultaneous face-body capture Application to simultaneous face-body capture

50. 50 Future directions Application to other domains Application to other domains More levels of hierarchy More levels of hierarchy Selection of multiple sub-regions Selection of multiple sub-regions Alternate system designs Alternate system designs

51. 51 Acknowledgements Prof. Pat Hanrahan, Prof. Brian Wandell, Prof. Chris Bregler, Prof. Gene Alexander, Prof. Marc Levoy, Cindy Chen, Ada Glucksman, Heather Gentner, Homan Igehy, Venkat Krishnamurthy, Tamara Munzner, François Guimbretière, Szymon Rusinkiewicz, Maneesh Agrawala, Lucas Pereira, Kari Pulli, Shorty, Sean Anderson, Reid Gershbein, Philipp Slusallek, Milton Chen, Mathew Eldridge, Natasha Gelfand, Olaf Hall - Holt, Humper, Brad Johanson, Sergey Brin, Dave Koller, John Owens, Kekoa Proudfoot, Kathy Pullen, Bill Mark, Dan Russel, Larry Page, Li-Yi Wei, Gordon Stoll, Julien Basch, Andrew Beers, Hector Garcia-Molina, Brian Freyburger, Mark Horowitz, Erika Chuang, Chase Garfinkle, John Gerth, Xie Feng, Craig Kolb, Toli, Mom, Dad, Holly Jones, Chris, Crystal, Lara, Grace Gamoso, Matt Hamre, Nancy Schaal, Aaron Jones, Bandit, Xiaoyuan Tu, Abigail, Shefali, Liza, Phil, Deborah, Brianna, Alejandra, Miss Dungan, Gabe, Sedona, Sharon, Gonzalo, and many other children whose names I can no longer remember