Can hog feature descriptor be used to create an efficient hash?

[KilianB]

The hog feature descriptor pools gradient vectors based on their unsigned direction. It is successfully used in pedestrian detection.

So far a derived descriptor based on https://lear.inrialpes.fr/people/triggs/pubs/Dalal-cvpr05.pdf is implemented but we are still left with way to many numbers to encode them into a short hash.

Images to compare	normalized hamming distance
Similar images
HQ - HQ:	0.000
HQ - LQ:	0.292
HQ - Copy:	0.119
HQ - Thumb	0.422
LQ - Copy	0.319
LQ - Thumb	0.396
Copy - Thumb	0.424
Unlike Images
HQ - Ballon:	0.496
HQ - Lena:	0.491
LQ - Ballon	0.484
LQ - Lena	0.483
Copy - Ballon	0.491
Copy - Lena	0.493
Thumb - Ballon	0.506
Thumb - Lena	0.481

While similar images can be differentiated from unlike images it's rather expensive, the hash is long. A todo on figuring out if the approach can be tweaked to produce usable results.

[KilianB]

5134663727920127093d2846a3c157dd3a8b4a12 doesn't look promising so far.