- image rep
- distribution of local intensity gradients can help characterize object appearance
key idea
-
Divide image window into cells (rectangle or radial)
-
each window sums up local 1-d histogram of gradient directions over pixels of the cell
- count frequency of gradients..?
-
normalization makes it invariant to illumination/shadows
-
captures object contour
limitation
diff from SIFT - local detector
- SIFT is for local keypoint matching
- normalized wrt dominant gradient (allows rotation invariant)
- HoG describes larger image regions
- normalized wrt neighborhood blocks (illumination invariant)