Minmax KL Learning + HOG on INRIA Person Database

1. Introduction

The code using less negative examples is available for download at minmax_hog_inria_concise.
The code using 12180 negative examples is available for download at minmax_hog_inria_12180.
There is another version producing exactly the same results but visulizing by itself without Matlab: minmax_hog_vis (63M).
The code using 48720 negative examples is available for download at minmax_hog_inria_48720 (118MB, run several hours).
There is another toy example for reweighting Generative Boost, C++ version and Matlab version. minmax_toy.

2. Results

I use varying number of features. And plot the ROC and DET. On ROC, I compute the AUC. On DET, I compute the EQ point, i.e., where false positive rate and miss rate are equal. The results reported in HOG paper (CVPR) is EQ 0.01.

EXP1

The images are 57x148, numbers of examples are as follows.

	Training	Testing
Positive	1540	564
Negative	12180	5047

For HOG, I extract 972 candidate features for each image.

# of features	AUC	EQ	Miss rate at false positive rate=1e-3	Miss rate at FPR=1e-4
50	0. 95827	0.102
100	0.9703	0.089
150	0.9728	0.085
200	0.9735	0.083
250	0.9753	0.079
300	0.9755	0.078
350	0.9763	0.077
400	0. 97517	0. 079	0. 74645	0.94858
500	0. 97693	0. 0728	0. 70567	0.93262
600	0.9773	0. 0735	0.71631	0.93972
700	0. 97702	0. 076	0. 72518	0. 9344
800	0. 9771	0.0759	0.69681	0.92908
900	0.9776	0.0727	0.69504	0.93617
950	0.9783	0.0727	0.70567	0.93972
Original HOG		0.01	0.04	0.1

When I flip the examples to get # of negatives double

		Training			Testing
Positive		1520x2=3080			564x2=1128
Negative		12180x2=24360			5047x2=10094

# of features	AUC		EQ	Miss rate at false positive rate=1e-3		Miss rate at FPR=1e-4
100	0.97885		0.0744	0.64184		0.81915
300	0.98337		0.066	0.52837		0.87234
500	0.98314		0.0637	0.60106		0.91844
650	0.98387		0.0620	0.62145		0.89982
700	0.98366		0.0624	0.59309		0.89362
750	0.98365		0.0629	0.58777		0.88475
800	0.98342		0.0613	0.60106		0.86613
850	0.98331		0.0629	0.63918		0.86702
900	0.98338		0.0629	0.63298		0.89982
972	0.9833		0.0623	0.61968		0.90603
Original HOG			0.01	0.04		0.1

When I re-cut the negative examples, together with flipping copies, there are 97440 negative examples for training.

0.9825

		Training			Testing
Positive		1540x2=3080			564x2=1128
Negative		48720x2=97440			20188x2=40376

# of features	AUC		EQ	Miss rate at false positive rate=1e-3		Miss rate at FPR=1e-4
100	0.97841		0.0730	0.47872		0.71809
300	0.98357		0.0638	0.41401		0.74025
500
650
700
750	0.0629	0.45656, 0.81294
800
850
900
972
Original HOG			0.01	0.04		0.1

3. TODO List

1. Visualize intermediate results.

2. Use correct training negative examples. Crop negative examples exhaustly --- negative examples mining.

3. Fit the histogram p_{n-1}(r_n) and f(r_n) using Gaussian or exponential.

4. Compare Generative Boosting + Gabor

5. Add shiftmax to HOG and Gabor to make them active

6. Combine Active Basis and HOG under Generative Boosting.

7. EM and mixture model

8. Add inhibition to simplify the reweighting, i.e., when updating histogram p_{n-1}(r_n) , independent features are not updated. This will improve the speed and make the generalization better.