Histogram of Oriented Gradients Rotation and transformation

Question

I am using HOG for detection of objects in my project. However when the object rotates the hog fails to correctly detect the object. Hence my confusion is that whether like SURF, SIFT, is HOG also rotation and transformation inVariance.

If not than I think I should perform affine transformation when computing HOG. Please advice me on this. Thank you.

Answer 1

HOG is not a rotation-scale invariant descriptor. Typically multi-scale detectors run the train detector at different image scales (as OpenCV detector does). For rotation invariance you can look at rotation invariant descriptors (SIFT,SURF,ORB,...). Alternatively you can simply include in your training set object appearing at various poses. However, this will include more complexity and uncertainty in the detector and depending on the objects you are trying to detect it might not be a viable solution.

Answer 2

What is the question? HOG is not rotation invariant feature. For rotation invariance use SIFT (RIFT) or rotate your image manually: find the main orientation of the descriptor, and assign that angle to the keypoint.