Feature Change
I modified the standard HOG features by adding a local normalization step.
After gradient magnitudes were computed, I took the whole gradient magnitude map
and convolved it (separably) with a Gaussian to obtain a local average of the grad mag.
Let v(p) be the grad magnitude at a pixel p. I computed v(p) / G*v(p) at each point and used
this value in the histogram instead of v(p). This particular process tends to accentuate
edges due to object boundaries over edges due to texture. It has been posited in the past
that a lot of HOG's descriptive power comes from these edges, as HOG on Sobel pre-processed
images can still perform well on some tasks.
Unfortunately, the sigma of said Gaussian has a lot to say about how well this works. I tried
to tune sigma with low values of STEP_SIZE and SCALE_STEP, since they ran quickly, but performance
was already saturated at these parameters... And the experiment takes a long time to run at discriminative
resolutions, so feature tuning is slow.
I didn't manage to significantly improve performance, but I by no means performed an exhaustive search.