I have Graduate Level- HOG implementation section in code/HOG folder. I wrote my own version DalalTriggs for HOG.
I directly implemented the multi scale face detection as it is very the same as single scale. Figure 1 shows a new training based on the threshold "-.3" and cell-size "6".
Figure 1a. Recall versus False-positives with cell-size 6 |
Figure 1b. Precision-recall comparison with cell-size 6 |
Figure 2. HOG - cell-size 4 |
Figure 3a. Recall versus False-positives with cell-size 4 |
Figure 3b. Precision-recall comparison with cell-size 4 |
Some results based on above parameters:
Dalal and Triggs method was leveraged for my HOG implementation. The Wikipedia implementation of Dalal and Triggs HOG feature cab be found in code/HOG. See "kamal_hog.m" and "hog_hist.m".
Figure 4a. Recall versus False-positives based on My Dalal and Triggs HOG implementation |
Figure 4b. Precision-recall comparison based on My Dalal and Triggs HOG implementation |
Figure 5a. Recall versus False-positives based on vl_feat Dalal and Triggs HOG implementation |
Figure 5b. Precision-recall comparison based on vl_feat Dalal and Triggs HOG implementation |
Some sample results: