The ideal amount of HoG features

Question

So, there are many options for how you can extract HoG functions. Using different orientations, a different number of pixels per cell and different block sizes.

But is there a standard or optimal configuration? I have 50x100 training images and I choose 8 orientation directions. I extract functions from training data to classify a car. But I really do not know what is "optimal."

For example, I have 2 configurations here, are there any reasons to choose one by one? Personally, I feel that the second option is the best choice, but why?

Answer 1

I used HOG to recognize the product. From what I understood at the time, you are pointing out the real problem with the standard HOG. There is simply no optimal configuration, it depends on the data set. If you have the optimal values ​​for your dataset, and then resize all the images in your dataset, you must also resize your values. Thus, for HOG there are no optimal values ​​"one size fits all".

But all is not lost. Instead, you should use a method that works "all the time." The idea is to make a spatial comparison of the pyramids . It just makes HOG on different scales and brings them together. A picture is worth a thousand words:

You can see that here level 2 is the standard HOG with a thin cell. But perhaps this is not the best scale (because the cells are too small and you just observe noise) (On the other hand, too large cells, for example, level 0, can be too large, and everywhere you will have uniform histograms). You can calculate the best weights for each level when you exercise in your data set, and you will know what are the optimal values, i.e. The most suitable cell size.