OpenCV HOGDescripter Python

Question

I was wondering if anyone knew why there is no documentation for HOGDescriptors in the Python bindings of OpenCV.

Maybe I\'ve just missed them, but the only code I\'

Answer 1

I was wondering the same. Almost none documentation can be found for OpenCV HOGDescriptor, other than the source cpp code.

Scikit-image has a good example page on extracting and illustrating HOG feature. It provides an alternative to explore HOG. It is documented here.

However, there is one thing to point out about scikit-image's hog implementation. Its Python code for hog function does not implement weighted vote for histogram orientation binning, but only does simple binning based on magnitude value falling into which bin. See its hog_histogram function. This is not following exactly Dalal and Triggs's paper.

Actually, I found that object detection based on OpenCV's implementation of HOG is more accurate than with the api from scikit-image. It makes sense to me, because weighted vote is important. By casting weighted votes to bins, variation in histogram is greatly reduced when gradient magnitude falls on or around the boundary. Chris McCormick wrote a very insightful blog on hog, in which orientation binning is clearly described as

For each gradient vector, it’s contribution to the histogram is given by the magnitude of the vector (so stronger gradients have a bigger impact on the histogram). We split the contribution between the two closest bins. So, for example, if a gradient vector has an angle of 85 degrees, then we add 1/4th of its magnitude to the bin centered at 70 degrees, and 3/4ths of its magnitude to the bin centered at 90.

I believe the intent of splitting the contribution is to minimize the problem of gradients which are right on the boundary between two bins. Otherwise, if a strong gradient was right on the edge of a bin, a slight change in the gradient angle (which nudges the gradient into the next bin) could have a strong impact on the histogram.

So, use OpenCV to compute hog if possible (haven't digged into its code and don't feel like doing so, but I suppose OpenCV's way of hog implementation is more appropriate). Not only I found an improvement in detection accuracy, but it also runs faster. Compared to scikit-image's hog code with wonderful comments, its documentation is almost none. Yet it is still feasible that one could get OpenCV's version working in practice - it's a matter of passing the right parameter for window size, cell size, block size, block stride, number of orientations, etc. Other parameters I just went with default.