P01 introduction cvpr2012 deep learning methods for vision
- 1. Deep Learning & Feature Learning Methods for Vision ’
- 3. Overview • – • – – – •
- 5. Motivation • • – •
- 6. What Limits Current Performance? • – •
- 7. Hand-Crafted Features • β – • – •
- 8. Mid-Level Representations • “ ” • •
- 9. Why Learn Features? • • – – – • – –
- 10. Why Hierarchy? fi
- 11. Hierarchies in Vision • – • –
- 13. Learning a Hierarchy of Feature Extractors • • • •
- 15. Classic Approach to Training • – – – • – –
- 18. Example Feature Learning Architectures
- 19. SIFT Descriptor
- 21. Filtering • –
- 22. Filtering • – – . . .
- 23. Translation Equivariance • – –
- 24. Filtering • – –
- 25. Filtering • – – – –
- 26. Normalization • •
- 27. Normalization • – –
- 28. Normalization • – –
- 29. Role of Normalization • – “ ” – – • |.|1 |.|1 |.|1 |.|1
- 30. Pooling • – – –
- 31. Role of Pooling • – –
- 32. Role of Pooling • • • •
- 33. Unsupervised Learning • • • – – –
- 34. Auto-Encoder
- 35. Auto-Encoder Example 1 • σ(WTz) σ(Wx) σ σ
- 36. Auto-Encoder Example 2 • Dz σ(Wx) σ
- 37. Auto-Encoder Example 2 • Dz σ(Wx) σ
- 38. Taxonomy of Approaches • – – – • – – • –
- 41. Information Flow in Vision Models • • – – • –
- 43. Why is Top-Down important? • • •
- 44. Multi-Scale Models • • • HOG Pyramid
- 45. Hierarchical Model • Input Image/ Features Input Image/ Features
- 46. Multi-scale vs Hierarchical Feature Pyramid Input Image/ Features
- 47. Structure Spectrum • – – – • – –
- 48. Structure Spectrum • – –
- 50. Structure Spectrum • – –
- 52. Structure Spectrum • – –
- 53. Structure Spectrum • – • – •
- 54. Structure Spectrum • – – –
- 55. Structure Spectrum • – – –
- 56. Performance of Deep Learning • • – • – • • – • – •
- 57. Summary • – • • •
- 58. Further Resources • • • • – •
Editor's Notes
- #11: All I am going to say about Neuroscience, although techniques do have strong connections.
- #14: Make clear that classic methods, e.g.convnets are purely supervised.
- #15: Need to bring outdiffereceswrt to existing ML stuff, mainly unsupervised learning part. Make use of unlabaled data (lots of it).
- #16: Restructure to bigger emphasis on unsupervised.Make clear that classic methods, e.g.convnets are purely supervised.
- #18: Winder and Brown paper. Slightly smoothed view of things.
- #19: Selection instead of normalization?
- #20: Note pooling is across space, not across Gabor channelNormalization is really nonlinear (small elements not rescaled)
- #21: Non-maximal suppression across VW. Like an L-InfnormalizationMax = k-means
- #32: Graph not clear. Explain better. Y-axis is change in value
- #33: Mention Leonardis & Fidler paper
- #34: Too far for labels to trickle down (vanishing gradients)Only information from layer below.Input is supervision.
- #37: Add overall energy
- #42: Not separate operations Do it at the same
- #43: Chriswilliams oral link
- #44: Occlusion mask: bootom right quad for sofa interpretationCan’t decide locally If you knew solution, would know what features to extract.
- #46: DPM is shape hierarchical HOG templates
- #47: DPM is shape hierarchical HOG templates
- #48: Song Chun ‘s clock