3. Object Classes Most recent work is at the object level
We perceive the world in terms of objects, belonging to different classes.
What are the differences between dogs and cats
What is common across different examples of objects in the class, shoes, trees…

4. We want in addition: 1. Individual Recognition4

5. 2. Object parts and sub-parts Called: Full Interpretation
Window
Mirror
Window
Door knob
Headlight
Back wheel
Bumper
Mirror, door knob are probably ‘query driven’
Front wheel
Headlight

6. 3. Action recognition (which 2 are different?)
2 here an non-drinking

7. 4. Agents Interactions 3 1 2 4 5 6

8. Class Non-class

9. Class Non-class

10. Is this an airplane?

11. Unsupervised Training Data

12. Features and Classifiers
In DNN -- the net produces features of the top layer
Previous work explored a broad range of features

13. Features used in the past: Generic Features
Simple (wavelets)
Complex (Geons)

14. Marr-Nishihara
Binford 1971, Generalized cylinders

15. Marr Net 2017 rotated versions of the object in the image
Last column – this are rotated versions of the object in the image
rotated versions of the object in the image

16. Past Class-specific Features: Common Fragments

17. Optima Features: Mutual Information I(C,F)
Class:
Feature:
I(F,C) = H(C) – H(C|F)

18. Star model Detected fragments ‘vote’ for the center location
Find location with maximal vote
In variations, a popular state-of-the art scheme

19. Hierarchies of sub-fragments (a ‘deep net’)
Detect the part itself by simpler sub-parts
Repeat at multiple levels, to obtain a hierarchy of parts and sub-parts

20. Example Hierarchies

21. Classification by Features Hierarchyx2 X1 X3 X4 X5
p(c,X,F) = p(c)Πp(xi|xi-) p(Fi|xi)

22. Global optimum can be found by max-sum message passing (two-pass computation)
X = argmax
[p(c,X,F) = p(c)Πp(xi|xi-) p(Fi|xi) ]

23. Context: Parts and Objects Results of two-pass computation
Outside = clips of the input image

24. Current use of probabilistic graph models

25. HoG Descriptor
Dallal, N & Triggs, B. Histograms of Oriented Gradients for Human Detection
What is (c) here? Mention Normalization
SIFT is similar, different details, multi-scale

26. SVM – linear separation in feature space

27. Optimal Separation Perceptron SVM
How do we actually find this optimal plane? We need to find a plane that maximizes the minimal separation.
Frank Rosenblatt, 1962 book, Principles of Neurodynamics.  
Perceptron
SVM
The Nature of Statistical Learning Theory, 1995
Rosenblatt, Principles of Neurodynamics 1962. 
Find a separating plane such that the closest points are as far as possible

28. The Margin +1 -1 Separating line: w ∙ x + b = 0
We can always call the separating line wx + b = 0, wx = 0 is the same plane, but through the origin. We can have the same plane with larger w and smaller b. We can choose a pair w,b that will make the further line be wx + b = 1
Calculating the margin. It is 2/|w|
Separating line: w ∙ x + b = 0
Far line: w ∙ x + b = +1
Their distance: w ∙ ∆x = +1
Separation: |∆x| = 1/|w|
Margin: 2/|w|

29. Using patches with HoG descriptors and classification by SVM
This is from Dalal-Triggs, HoG + SVM. Felzenszwalb extends this: also uses HoG, the description is more complex and includes parts and their locations. What is the Phi here? I think you simply include the coefficient b, since we only have here wf without b?
Using patches with HoG descriptors and classification by SVM
Person model: HoG

30. DPM: Adding Parts
In the object-model, the Fi is a vector of coefficients for part-i, learned by SVM during training. The vi is 2 numbers, si, a scale, is 1 number, the ai is 2 numbers, bi is 2 numbers.

31. Bicycle model: root, parts, spatial map
A bicycle model, and a person. the root, parts, and spatial map. The Filter in the model is the vector of SVM weights. The figure shows the orientations that have positive w values.
Person model

32. Deep Learning

34. ImageNet

37. AlexNet

38. On the history deep learning

39. A Neural Network Model A network of ‘neurons’ with multiple layers
Repeating structure, linear, non-linear
Automatic learning of weights between units

40. The McCulloch–Pitts neuron (1943)
Relu

41. Perceptron learning
yj = f(xj)

42. Back-propagation 1986

43. LeNet 1998
Essentially the same as the current generation

44. MNIST data set

45. Hinton Trends in Cognitive Science 2007
The goal: unsupervised
Restricted Boltzmann Machines
Combining generative model and inference
CNN are feed-forward and massively supervised

46. Basic structure of deep nets.
Not detailed here, but make sure you know the layers structure and repeating 3-layer arrangement