Is there any non-naive approach to sliding window algorithm?

Question

I'm trying to extract features out of a 2D rgb image (and later to be extended to 3D depth data).

Basically what I am doing is to slide a 64 x 128 window around my image (and later on, on smaller scale version of the image), with stride of 8x8. However I find that this approach is really really slow.

Is there any sliding window algorithm that could speed up this process ?

Edit: I'm performing detection by extracting HoG features (and its 3d variants) from the sliding window, before using SVM to predict the result.

Answer 1

If you're using the sliding window only for detection of the object, then this paper is quite a good reference for a branch and bound approach with scale invariance.