Histogram oriented gradients and face orientation histograms

Question

I do not quite understand the difference between HOG and EOH. Hog based on derived images EOH based on edge directions. It seems that the HOG also somehow represents the EOH.

Could you give me some explanation about how EOH differs from HOG and its advantages over EOH compared to HOG. In what circumstances can we use EOH compared to HOG?

Answer 1

I think that the main difference is that for HOG the actual direction of the gradient is calculated and then bind, where for EOH the edge orientation is estimated by finding the maximum answer for the set of cores of the extreme filter. So you can say that HOG binders after calculating the gradient, where EOH directly calculates the gradient in the bins. Depending on the number of bins you want, each will be faster than the other.

In EOH, light and dark light edges are usually treated the same, and therefore orientations are in the range from 0 to pi, where in HOG bins usually cover the full 2 ​​* pi. You can easily do EOH do this too.

Answer 2

Although, I don’t think there is a unique definition for them, and the details can vary at different levels (for example, how to calculate the orientation (or orientation gradients)), the key is that the histograms constructed by the histograms of the edge orientation take into account only the gradient of pixels corresponding to edges (which, in turn, are calculated by some other method, for example, canny edges), while histograms of oriented gradients take into account all the gradients of each pixel.

Maybe I'm wrong, but that is how I implemented EOH: http://robertour.com/2012/01/26/edge-orientation-histograms-in-global-and-local-features/