From Unsupervised to Semi-Supervised Event Detection

1. From Unsupervised to Semi-Supervised Event Detection Wen-Sheng Chu Robotics Institute, Carnegie Mellon University July 9, 2013 1 Jeffery CohnFernando De la Torre

2. Outline 1. Unsupervised Temporal Commonality Discovery (Chu et al, ECCV12) 2. Personalized Facial Action Unit Detection (Chu et al, CVPR13) 2

3. Unsupervised Commonality Discovery in Images Where are the repeated patterns? 3 (Chu10, Mukherjee11, Collins12)

4. Unsupervised Commonality Discovery in Videos? We name it Temporal Commonality Discovery (TCD). Goal: Given two videos, discover common events in an unsupervised fashion. 4

5. TCD is hard! 1) No prior knowledge on commonalities We do not know what, where and how many commonalities exist in the video 2) Exhaustive search are computationally prohibitive E.g., two videos with 300 frames have >8,000,000,000 possible matches. possible locations possible lengths possibilities/sequence Another possibilities/sequence 5

6. Formulation 6 Integer programming!

7. Optimization: Interpretation 7

8. Optimization: Native Search Complexity 8

9. Optimization: Branch-and-Bound Similar to the idea of ESS (Lampert08), we search the space by splitting intervals. 9

10. Optimization: Branch-and-Bound Bounding histogram bins 10

11. 1. Bounding L1 distance: 2. Intersection similarity: 3. X2 distance: Optimization: Branch-and-Bound 11

12. Unlikely search regions (B1,E1,B2,E2; -10) Searching Structure (B1,E1,B2,E2; 32) Priority queue (sorted by bound scores) (B1,E1,B2,E2; -50) (B1,E1,B2,E2; -105) State S = (Rectangle set; score) 12

13. (B1,E1,B2,E2; -105) Algorithm (B1,E1,B2,E2; 32) Priority queue (sorted by bound scores) (B1,E1,B2,E2; -50) (B1,E1,B2,E2; -105) Top state 1. Pop out the top state 2. Split 13

14. (B1,E1,B2,E2; -105) Algorithm (B1,E1,B2,E2; 32) Priority queue (sorted by bound scores) (B1,E1,B2,E2; -50) Top state (B1,E1,B2,E2; -76) (B1,E1,B2,E2; -61) 3. Compute bounding scores 4. Push back the split states 14

15. Algorithm (B1,E1,B2,E2; 32) Priority queue (sorted by bound scores) (B1,E1,B2,E2; -50) Top state (B1,E1,B2,E2; -76) (B1,E1,B2,E2; -61) The algorithm stop when the top state contains an unique rectangle. Omit most of the search space with large distances 15

16. Compare with Relevant Work 1. Difference between TCD and ESS [1]/STBB[2] Different learning framework: Unsupervised v.s. Supervised New bounding functions for TCD 2. Difference between TCD and [3] Different objective: Commonality Discovery v.s. Temporal Clustering [1] Efficient subwindow search: A branch and bound framework for object localization, PAMI 2009. [2] Discriminative video pattern search for efficient action detection, PAMI 2011. 16

17. Experiment (1): Synthesized Sequence Histograms of the discovered pair of subsequences 17

18. Experiment (2): Discover Common Facial Actions RU-FACS dataset* Interview videos with 29 subjects 5000~8000 frames/video Collect 100 segments that containing smiley mouths (AU- 12) Evaluate in terms of averaged precision 18 * Automatic recognition of facial actions in spontaneous expressions, Journal of Multimedia 2006.

19. Experiment (2): Discover Common Facial Actions 19

20. Parametric settings for Sliding Windows (SW) Log of #evaluations: Quality of discovered patterns: a Experiment (2): Speed Evaluation Speed #evaluation of the distance function log nT C D nSW i d(r SW i ) d(r T C D ) 20

21. Experiment (2): Discover Common Facial Actions Compare with LCCS* on -distance 21 * Frame-level temporal calibration of unsynchronized cameras by using Longest Consecutive Common Subsequence, ICASSP 2009.

22. Experiment (3): Discover Multiple Common Human Motions CMU-Mocap dataset: http://mocap.cs.cmu.edu/ 15 sequences from Subject 86 1200~2600 frames and up to 10 actions/seq Exclude the comparison with SW because it needs >1012 evaluations 22

23. Experiment (3): Discover Multiple Common Human Motions 23

24. Experiment (3): Discover Multiple Common Human Motions Compare with LCCS* on -distance 24

25. Extension: Video Indexing Goal: Given a query , find the best common subsequence in the target video A straightforward extension: Temporal Search Space 25

26. A Prototype for Video Indexing 26

27. Summary 27

28. Questions? [1+ Common Visual Pattern Discovery via Spatially Coherent Correspondences, In CVPR 2010. [2+ MOMI-cosegmentation: simultaneous segmentation of multiple objects among multiple images, In ACCV 2010. [3+ Scale invariant cosegmentation for image groups, In CVPR 2011. [4+ Random walks based multi-image segmentation: Quasiconvexity results and GPU-based solutions, In CVPR 2012. [5+ Frame-level temporal calibration of unsynchronized cameras by using Longest Consecutive Common Subsequence, In ICASSP 2009. [6+ Efficient ESS with submodular score functions, In CVPR 2011. 28 http://humansensing.cs.cmu.edu/wschu/

29. Outline 1. Unsupervised Temporal Commonality Discovery (Chu et al, ECCV12) 2. Selective Transfer Machine for Personalized Facial Action Unit Detection (Chu et al, CVPR13) 29

30. AU 6+12 Facial Action Units (AU) 30

31. Main Idea 31

32. Related Work: Features 32

33. Related Work: Classifiers 33

34. Feature Bias Person specific! 34

35. Occurrence Bias 35

36. Selective Transfer Machine (STM) Formulation Maximizes margin of penalized SVM Minimize distribution mismatch 36

37. Goal (1): Maximize penalized SVM margin margin penalized loss 37

38. Goal (2): Minimize Distribution Mismatch Kernel Mean Matching (KMM)* 38 * Covariate shift by kernel mean matching, Dataset shift in machine learning, 2009.

39. Goal (2): Minimize Distribution Mismatch Groundtruth Bad estimator for testing data! 39

40. Better fitting! Groundtruth Selection by reweighting training data 40 Goal (2): Minimize Distribution Mismatch

41. 41

42. 42 Optimization: Alternate Convex Search

43. 43 Optimization: Alternative Convex Search

44. Compare with Relevant Work 44 [1] "Covariate shift by kernel mean matching," Dataset shift in machine learning, 2009. [2] "Transductive inference for text classification using support vector machines," In ICML 1999. [3] "Domain adaptation problems: A DASVM classification technique and a circular validation strategy," PAMI 2010.

45. Experiments Features SIFT descriptors on 49 facial landmarks Preserve 98% energy using PCA 45 Datasets #Subjects #Videos #Frm/vid Content CK+ 123 593 ~20 NeutralPeak GEMEP-FERA 7 87 20~60 Acting RU-FACS 29 29 5000~7500 Interview

46. Experiment (1): Synthetic Data 46

47. Two protocols PS1: train/test are separate data of the same subject PS2: training subjects include test subject (same protocol in [2]) GEMEP-FERA Experiment (2): Comparison with Person- specific (PS) Classifiers 47

48. Experiment (2): Selection Ability of STM 48

49. 123 subjects, 597 videos, ~20 frames/video Experiment (3): CK+ 49

50. Experiment (4): GEMEP-FERA 50 7 subjects, 87 videos, 20~60 frames/video

51. 29 subjects, 29 videos, 5000~7000 frames/vid Experiment (5): RU-FACS 51

52. Summary Person-specific biases exist among face- related problems, esp. facial expression We propose to alleviate the biases by personalizing classifiers using STM Next Joint optimization in terms of Reduce the memory cost using SMO Explore more potential biases in face problems, e.g., occurrence bias 52

53. Questions? [1] "Covariate shift by kernel mean matching," Dataset shift in machine learning, 2009. [2] "Transductive inference for text classification using support vector machines," In ICML 1999. [3] "Domain adaptation problems: A DASVM classification technique and a circular validation strategy," PAMI 2010. *4+ Integrating structured biological data by kernel maximum mean discrepancy, Bioinformatics 2006. *5+ Meta-analysis of the first facial expression recognition challenge, IEEE Trans. on Systems, Man, and Cybernetics, Part B, 2012. 53 http://humansensing.cs.cmu.edu/wschu/

From Unsupervised to Semi-Supervised Event Detection

Technology