These are chat archives for menpo/menpo

5th
Jun 2014
Epameinondas Antonakos
@nontas
Jun 05 2014 16:52
Hey @dubzzz ! Sorry for the late reply...

Regarding your HOGs-related questions:
1) In general, a rational value for cellHeightAndWidthInPixels parameter would be an integer in the region [2,16]. However, we mostly use the default value, which is 8. I think you should focus on that case.

2) Regarding your HOG::ZhuRamananHOGdescriptor at line 87 in HOG.cppcomment: this code is not written by me; it's Zhu & Ramanan's code as noted in thelicense_ZhuRamananHOGdescriptor.txt` file. So I've never actually went through the code and I don't really recommend to change it. However, your comment is indeed interesting and we could discuss it in person if you want.

3) The same stands for your DalalTriggsHOGdescriptor question. The code again is not mine, however your observation is totally correct. In order to solve this, we can consider as a fact that blockNorm = sqrt(blockNorm) will only return a number greater or equal to zero. Thus, if blockNorm > 0, the block[i][j][k] should take the value it takes now. If blockNorm == 0 (else case), block[i][j][k] will be infinite and then it must be clipped to the l2normClipping parameter value. So from line 425:

if (blockNorm > 0) {
    block[i][j][k] = h[y+i][x+j][k] / blockNorm;
    if (block[i][j][k] > l2normClipping)
        block[i][j][k] = l2normClipping;
}
else
    block[i][j][k] = l2normClipping;

or something similar.

These were good remarks Nicolas. I will be in the meeting on Monday, so we can further discuss them and we should consider making the changes to the original menpo code.