Gender recognition from frontal facial images within project FP7-ICT-247525-HUMAVIPS

Tunning parammeter C of SVM OCAS solver and other parameters of HOG descriptor

<< zpět

Overview

This experiment has same goals as experiment 01, but here will be used HOG (Histogram Of Gradients) descirptor for representation of input images.

In first part will be tested with simple HOG. Next part contain results from classifiers based on extended descriptor by its square.

Results

Parameters of experiment
imSize [h x w]80 x 64
winSize [h x w]60 x 40
win positioncentered

SQUARED HOG, nBins = 9, cellSize = 2, blockSize = 10:
CTraining errorValidation error
0.00100022.73%24.45%
0.010000 9.62%11.86%
0.100000 4.24%11.14%
1.000000 0.54%12.78%

SQUARED HOG, nBins = 9, cellSize = 2, blockSize = 20:
CTraining errorValidation error
0.00100036.89%38.91%
0.01000019.92%21.28%
0.100000 9.63%12.27%
1.000000 4.38%12.27%

HOG, nBins = 9, cellSize = 2, blockSize = 4:
CTraining errorValidation error
0.00100010.10%12.97%
0.010000 4.47%11.46%
0.100000 0.51%12.85%
1.000000 0.00%14.03%

SQUARED HOG, nBins = 9, cellSize = 2, blockSize = 4:
CTraining errorValidation error
0.001000 9.92%12.95%
0.010000 4.07%11.42%
0.100000 0.24%13.04%
1.000000 0.00%13.32%

SQUARED HOG, nBins = 12, cellSize = 2, blockSize = 4:
CTraining errorValidation error
0.001000 9.86%12.97%
0.010000 3.53%11.51%
0.100000 0.05%13.53%
1.000000 0.00%13.86%

SQUARED HOG, nBins = 12, cellSize = 4, blockSize = 4:
CTraining errorValidation error
0.00100015.12%17.04%
0.010000 9.82%12.88%
0.100000 6.50%12.87%
1.000000 3.80%13.63%

Testing error for minimal validation error: 25.51%

Resources

[1] Corinna Cortes, Vladimir Vapnik: Support-vector networks - http://www.springerlink.com/content/k238jx04hm87j80g/
[2] Vojtech Franc, Soeren Sonnenburg: LIBOCAS - Library implementing OCAS solver for training linear SVM classifiers from large-scale data - http://cmp.felk.cvut.cz/~xfrancv/ocas/html/
<< zpět