CS4495/6495 Introduction to Computer Vision. 8C-L3 Support Vector Machines

Transcription

1 CS4495/6495 Introduction to Computer Vision 8C-L3 Support Vector Machines

2 Discriminative classifiers Discriminative classifiers find a division (surface) in feature space that separates the classes Several methods Nearest neighbors Boosting Support Vector Machines

3 Linear classifiers

4 Linear classifiers

5 Lines in R 2 Let w p q x x y px qy b 0

6 Lines in R 2 x, y 0 0 D w Let p w q x x y px qy b 0 w x b 0 Distance from point to line D px qy b 0 0 p q 2 2 wx w b

7 Linear classifiers Find linear function to separate positive and negative examples xi positive : xi w b 0 x negative : x w b 0 i i Which line is best?

8 Support Vector Machines (SVMs) Discriminative classifier based on optimal separating line (2D case) Maximize the margin between the positive and negative training examples

9 How to maximize the margin? x positive ( y 1) : x w b 1 i i i x negative ( y 1) : x w b 1 i i i For support vectors, x i w b 1 x Dist. between point and line: i wb w wxb 1 For support vectors: w w support vectors C. Burges, 1998 margin M M w w w

10 Finding the maximum margin line 1. Maximize margin 2 w 2. Correctly classify all training data points: x positive ( y 1) : x w b 1 i i i x negative ( y 1) : x w b 1 i i i 3. Quadratic optimization problem: C. Burges, T Minimize ww 2 Subject to y ( x w ) 1 i i b

11 Finding the maximum margin line Solution: w i yixi i learned weight support vector The weights α i are non-zero only at support vectors C. Burges, 1998

12 Finding the maximum margin line Solution: w i yixi b y i w x i i i i i i w x b y x x b Classification function: f( x) sign w x b sign i ixi x b (for any support vector) if f(x) < 0, classify as negative if f(x) > 0, classify as positive C. Burges, 1998 Dot product only!

13 Questions What if the features are not 2d? What if the data is not linearly separable? What if we have more than just two categories?

14 Questions What if the features are not 2d? What if the data is not linearly separable? What if we have more than just two categories?

15 Questions What if the features are not 2d? Generalizes to d-dimensions Replace line with hyperplane

16 Person detection with HoG s & linear SVM s Map each grid cell in the input window to a histogram counting the gradients per orientation Train a linear SVM using training set of pedestrian vs. nonpedestrian windows Dalal and Triggs, CVPR 2005 Code:

17

18 Questions What if the features are not 2d? What if the data is not linearly separable? What if we have more than just two categories?

19 Non-linear SVMs Datasets that are linearly separable with some noise work out great: 0 x

20 Non-linear SVMs But what are we going to do if the dataset is just too hard? 0 x

21 Non-linear SVMs How about mapping data to a higher-dimensional space: 0 x x 2 0 x

22 Non-linear SVMs How about mapping data to a higher-dimensional space: 0 x x 2 0 x

23 Non-linear SVMs: Feature spaces General idea: Original input space can be mapped to some higher-dimensional feature space Φ: x φ(x) Andrew Moore where the training set is easily separable

24 The kernel trick We saw linear classifier relies on dot product between vectors. Define: K x i, x j = x i x j = x i T x j

25 The kernel trick If every data point is mapped into high-dimensional space via some transformation Φ: x φ x the dot product becomes: K x i, x j = φ x i T φ x j A kernel function is a similarity function that corresponds to an inner product in some expanded feature space

26 E.g., 2D vectors x = [x 1 x 2 ] Let K x i, x j = 1 + x i T x j 2 Need to show that K x i, x j = φ x i T φ x j : K x i, x j = 1 + x T 2 i x j = 1 + x i1 2 x j x i1 x j1 x i2 x j2 + x i2 2 x j x xi1 x j1 + 2x i2 x j2 = 1 x i1 2 1 x j1 2 = φ x i T φ x j 2x i1 x i2 x i2 2 2x i1 2x j1 x j2 x j2 2 2x i2 T 2x j1 2x j2 where φ x = 1 x 1 2 2x 1 x 2 x 2 2 2x 1 2x 2 Andrew Moore

27 Nonlinear SVMs The kernel trick: Instead of explicitly computing the lifting transformation φ x, define a kernel function K: K x i, x j = φ x i T φ x j This gives a nonlinear decision boundary in the original feature space: i y ( x T x) b y K( x, x) b i i i i i i i

28 Examples of kernel functions Linear x T x i j i j K( x, x ) Number of dimensions: N (just size of x)

29 Examples of kernel functions Gaussian RBF K ( xi, x j) exp x i x 2 2 j 2 Number of dimensions: Infinite j 1 2 ( xx) exp x x exp x exp x 2 j! j0

30 Examples of kernel functions Histogram intersection K( x, x ) min( x ( k), x ( k)) i j i j k Number of dimensions: Large. See: Alexander C. Berg, Jitendra Malik, "Classification using intersection kernel support vector machines is efficient", CVPR, 2008

31 SVMs for recognition: Training 1. Define your representation 2. Select a kernel function 3. Compute pairwise kernel values between labeled examples 4. Use this kernel matrix to solve for SVM support vectors & weights

32 SVMs for recognition: Prediction To classify a new example: 1. Compute kernel values between new input and support vectors 2. Apply weights 3. Check sign of output

33 Learning gender with SVMs Learning Gender with Support Faces [Moghaddam and Yang, TPAMI 2002]

34 Face alignment processing Processed faces

35 Face gender classification by SVM Training examples: 1044 males, 713 females Images reduced to 21x12 (!!) Experiment with various kernels, select Gaussian RBF 2 xi x j K ( xi, x j) exp 2 2

36 Face gender classification by SVM Training examples: 1044 males, 713 females Images reduced to 21x12 (!!) Experiment with various kernels, select Gaussian RBF 2 x x j K ( x, x j ) exp 2 2

37 Support Faces

38 Moghaddam and Yang: Gender Classifier Performance

39 Gender perception experiment: How well can humans do? Subjects: 30 people (22 male, 8 female), ages mid-20 s to mid-40 s Test data: 254 face images (60% males, 40% females) Low res and high res versions Task: Classify as male or female, forced choice No time limit Moghaddam and Yang, Face & Gesture 2000

40 Human Performance Error Error

41 Careful how you do things?

42 Careful how you do things?

43 Human vs. Machine SVMs performed better than any single human test subject, at either resolution

44 Hardest examples for humans Moghaddam and Yang, Face & Gesture 2000

45 Questions What if the features are not 2d? What if the data is not linearly separable? What if we have more than just two categories?

46 Multi-class SVMs Combine a number of binary classifiers One vs. all Training: Learn an SVM for each class vs. the rest Testing: Apply each SVM to test example and assign to it the class of the SVM that returns the highest decision value One vs. one Training: Learn an SVM for each pair of classes Testing: Each learned SVM votes for a class to be assigned to the test example

47 SVMs: Pros and cons Pros Many publicly available SVM packages: Kernel-based framework is very powerful, flexible Often a sparse set of support vectors compact when testing Works very well in practice, even with very small training sample sizes Adapted from Lana Lazebnik

48 SVMs: Pros and cons Cons No direct multi-class SVM, must combine two-class SVMs Can be tricky to select best kernel function for a problem Nobody writes their own SVMs Computation, memory During training time, must compute matrix of kernel values for every pair of examples Learning can take a very long time for large-scale problems Adapted from Lana Lazebnik