Ensemble of Exemplar-SVMs for Object Detection and Beyond

Transcript

1. Ensemble of Exemplar- SVMs for Object Detection and Beyond Tomasz Malisiewicz, Abhinav Gupta and Alexei A. Efros ICCV, 2011
2. Abstract
3. Abstract • Object Detection
4. Abstract • Object Detection • On par with state of the art
5. Abstract • Object Detection • On par with state of the art • Much simpler
6. Abstract • Object Detection • On par with state of the art • Much simpler • At only a modest computational cost
7. Abstract • Object Detection • On par with state of the art • Much simpler • At only a modest computational cost • Central benefit: explicit association between each detection and one training example
8. Motivation
9. Motivation • Common Computer Vision tasks:
10. Motivation • Common Computer Vision tasks: • Image classification
11. Motivation • Common Computer Vision tasks: • Image classification • Object detection
12. Motivation • Common Computer Vision tasks: • Image classification • Object detection • bounding box
13. Motivation - Object Detection • Can we reason with bounding box? BUS
14. Motivation - How can we reason?
15. Motivation - How can we reason? • Obtain Association with a very similar exemplar from training
16. Motivation - How can we reason? • Obtain Association with a very similar exemplar from training • This is what mind does
17. Motivation - How can we reason? • Obtain Association with a very similar exemplar from training • This is what mind does • Enough data is currently available
18. Motivation - How can we reason? • Obtain Association with a very similar exemplar from training • This is what mind does • Enough data is currently available • Any kind of meta data could be transferred
19. Exemplars
20. Motivation - Exemplar Theory
21. Motivation - Exemplar Theory • Associating a new instance with something seen in the past
22. Motivation - Exemplar Theory • Associating a new instance with something seen in the past • Exemplar theory in cognitive psychology
23. Motivation - Exemplar Theory • Associating a new instance with something seen in the past • Exemplar theory in cognitive psychology • Case-based reasoning in AI
24. Motivation - Exemplar Theory • Associating a new instance with something seen in the past • Exemplar theory in cognitive psychology • Case-based reasoning in AI • Instance-based learning in ML
25. Exemplar Reasoning is Non-parametric
26. Exemplar Reasoning is Non-parametric KNN: non-parametric
27. Exemplar Reasoning is Non-parametric KNN: non-parametric SVM: parametric
28. Exemplar Theory in Computer Vision
29. Exemplar Theory in Computer Vision • Object Alignment • Scene Recognition • Image Parsing • Object Detection (not successful)
30. Non-parametric Object Detection
31. Non-parametric Object Detection • has not been competitive against discriminative approaches
32. Non-parametric Object Detection • has not been competitive against discriminative approaches • Why?
33. Non-parametric Object Detection • has not been competitive against discriminative approaches • Why? • Massive Amount of Negative data
34. Non-parametric Object Detection • has not been competitive against discriminative approaches • Why? • Massive Amount of Negative data • Classification vs Detection and KNN
35. Motivation - Negative Data
36. Motivation - Negative Data • Non-parametric methods are not suitable
37. Motivation - Negative Data • Non-parametric methods are not suitable • Parametric methods handle large amount of negative data very well
38. Motivation - Negative Data • Non-parametric methods are not suitable • Parametric methods handle large amount of negative data very well • HOG
39. Motivation - Negative Data • Non-parametric methods are not suitable • Parametric methods handle large amount of negative data very well • HOG • DPM
40. Motivation - Negative Data • Non-parametric methods are not suitable • Parametric methods handle large amount of negative data very well • HOG • DPM
41. Motivation - Negative Data
42. Motivation - Negative Data • SVM can handle negative data parametrically
43. Motivation - Negative Data • SVM can handle negative data parametrically • No negative data is stored (vs KNN)
44. Motivation - Negative Data • SVM can handle negative data parametrically • No negative data is stored (vs KNN) • Used by HOG
45. Parametric Approach
46. Parametric Approach • Very good representation of negative data
47. Parametric Approach • Very good representation of negative data • What about positive data?
48. Parametric Approach • Very good representation of negative data • What about positive data? • implicit assumption that all positive examples are visually related
50. Parametric Approach • Very good representation of negative data • What about positive data? • implicit assumption that all positive examples are visually related • results in over generalized models
51. Desirable Approach
52. Desirable Approach • All strengths of HOG/DPM
53. Desirable Approach • All strengths of HOG/DPM • powerful descriptor
54. Desirable Approach • All strengths of HOG/DPM • powerful descriptor • discriminative framework
55. Desirable Approach • All strengths of HOG/DPM • powerful descriptor • discriminative framework • handle massive amount of negatives
56. Desirable Approach • All strengths of HOG/DPM • powerful descriptor • discriminative framework • handle massive amount of negatives • Not rigidly representing positives
57. Desirable Approach • All strengths of HOG/DPM • powerful descriptor • discriminative framework • handle massive amount of negatives • Not rigidly representing positives • Good Association for meta-data transfer
58. Desirable Approach • All strengths of HOG/DPM • powerful descriptor • discriminative framework • handle massive amount of negatives • Not rigidly representing positives • Good Association for meta-data transfer Parametric Negatives Non-parametric Positives
59. Exemplar-SVMs • Learn a model for each positive example • HOG features • linear SVM classifier
60. Exemplar-SVMs • Learn a model for each positive example • HOG features • linear SVM classifier
61. Exemplar-SVMs • Training • Single Positive example • Millions of negative examples (sliding windows) - from images not containing any in-class instances
62. Large Scale Training • Use parallel Training on clusters
63. Exemplar-SVMs • Testing • Each sliding window is given to all Exemplar-SVMs • Highest score is the detection
64. Qualitative Examples
72. Meta-Data Transfer
78. Thank You Any Questions?