update hog

[hubutui]

1. remove unneeded function declaration
2. specify memoryview order
3. image is row-major by default
4. add visualize optional flag, return a HOG image if True
5. update document
6. update hog test, use face image from scipy
7. add channel_first optional flag, specify if the input array is in
   channel-first format

[hubutui]

@patricksnape Hey, could you check if this pr could be merged?

[patricksnape]

Can we fix the RGB before merging?

[patricksnape]

I guess this is because ignoring memory order is incorrect

[hubutui]

I create a simple demo in C++ using the C library to calculate HOG features, see https://github.com/hubutui/vlfeat-hog-example. The extracted feature is saved to a txt file one value per line with no reshape operation. So we could load it later on with np.loadtxt, reshape it to [dim, h, w], where dim is HOG feature dimension, its value is 31 for the default UoCTTI variant. Then we swap its axes to [h, w, dim] as we used in cyvlfeat. numpy.testing.assert_allclose shows that the HOG feauture calculated from C API and cyvlfeat are close enough for gray scale image, but not for RGB color image.

[patricksnape]

The only other thing I would request is a flag to be able to control this behaviour - in case the user wants to pass something where the channels are already first. Maybe something like channels_first=False?

[patricksnape]

This is amazing work! The mat files are really really good tests and you approached this really methodically which makes it really easy for me to review.

I had one annoying nitpick comment and then we are good to go and I can make a release.

Thanks a lot for this work!

[hubutui]

@patricksnape OK, I have update again. I think we could merge it now. I'd like to update sift before we could make a new release.

[patricksnape]

Thanks for your hard work - this was an excellent PR!