Presentation is loading. Please wait.

Presentation is loading. Please wait.

Kullback-Leibler Boosting Ce Liu, Hueng-Yeung Shum Microsoft Research Asia CVPR 2003 Presented by Derek Hoiem.

Similar presentations


Presentation on theme: "Kullback-Leibler Boosting Ce Liu, Hueng-Yeung Shum Microsoft Research Asia CVPR 2003 Presented by Derek Hoiem."— Presentation transcript:

1 Kullback-Leibler Boosting Ce Liu, Hueng-Yeung Shum Microsoft Research Asia CVPR 2003 Presented by Derek Hoiem

2 RealBoost Review Start with some candidate feature set Initialize training sample weights Loop:  Add feature to minimize error bound  Reweight training examples, giving more weight to misclassified examples  Assign weight to weak classifier according to weighted error of training samples  Exit loop after N features have been added

3 The Basic Idea of KLBoosting Similar to RealBoost except:  Features are general linear projections  Generates optimal features  Uses KL divergence to select features  Finer tuning on coefficients

4 Linear Features KLBoosting: VJ Adaboost:

5 What makes a feature good? KLBoosting: RealBoost:  Minimize upper bound on classification error

6 Creating the feature set Sequential 1-D Optimization  Begin with large initial set of features (linear projections)  Choose top L features according to KL-Div  Initial feature = weighted sum of L features  Search for optimal feature in directions of L features

7 Example Initial feature set: x x x x x x x x

8 Example Top two features (by KL-Div): x x x x x x x x w1w1 w2w2

9 Example Initial feature (weighted combo by KL): x x x x x x x x w1w1 w2w2 f0f0

10 Example Optimize over w 1 x x x x x x x x w1w1 w2w2 f1f1 f 1 = f 0 + B * w 1 B = -a 1..a 1

11 Example Optimize over w 2 x x x x x x x x w1w1 w2w2 f2f2 f 2 = f 1 + B * w 2 B = -a 2..a 2 (and repeat…)

12 Creating the feature set First three features Selecting the first feature

13 Creating feature set

14 Classification = ½ in RealBoost

15 Parameter Learning With each added feature k:  Set first a 1..a k-1 to current optimal value  Set a k to 0  Minimize recognition error on training:  Solve using greedy algorithm

16 KLBoost vs AdaBoost 1024 candidate features for AdaBoost

17 Face detection: candidate features 52,400  2,800  450

18 Face detection: training samples 8760 faces + mirror images 2484 non-face images  1.34B patches Cascaded classifier allows bootstrapping

19 Face detection: final features top ten global semantic global not semantic local

20 Results x x x x 885853 Schneiderman (2003) Test time:.4 sec per 320x240 image

21 Comments Training time? Which improves performance:  Generating optimal features?  KL feature selection?  Optimizing alpha coefficients?


Download ppt "Kullback-Leibler Boosting Ce Liu, Hueng-Yeung Shum Microsoft Research Asia CVPR 2003 Presented by Derek Hoiem."

Similar presentations


Ads by Google