Are the HOG features of the images good enough to do image classification alone? Why are their variances low?

Question

I've also asked this question here:https://cs.stackexchange.com/questions/77884/are-the-hog-features-of-the-images-good-enough-to-do-image-classification-alone

I've been working on classifying human poses like the pictures below. For that, I'm computing HOG features, and while in MATLAB, after using the command extractHOGfeatures, it's giving me 1764 HOG features, I also notice that maximum variance of all the 1764 HOG features are as low as 0.04 for a sample size of 2660, meaning the HOG features aren't probably changing much from sample-to-sample. My questions are:

1) why do you think we're getting low variances for HOG features? 2) what could we intuitively say about HOG features of these pose images? Are they good or interesting enough to classify images alone?

Answer 1

Definitely not sufficient to detect;

First HoG is not scale invariant - you change resize your image and handle this problem-.

Second HoG is not rotation invariant - you should include vary poses of action but this makes your solution more complex-

In my opinion; you should choose another kind of feature descriptors which can handle rotation and scale invariance. (SIFT, SURF, ORB ...)

Maybe you can use spatio-temporal information for this problem - It will be very useful-

Thanks.

And one more suggestion is about Neural Networks. If you have enough dataset for your problem, you can use "Neural Networks" which can be more roboust to other kind of feature based methods.