In order to extract texture-based features; LBPs (Local Binary Patterns) are used in many background-foreground detection algorithms. It is easy to understand why those features are tolerant against the illumination. However; there is one of their structural properties that i could not comprehend. What is the exact reason for using a "circular mask" rather than a "rectangular one", when we compute the pattern? For instance, in this paper, they claimed that their modification is better than ordinary LBP; though they did not specify the reason why.
Well first, i thought it is for obtaining a pattern which is rotation invariant; but... This is just my assumption, of course.
Does anyone know why circular structure is better than rectangular one?
A key feature of the circle is that every point on the circumference has the same distance from the center. If you want to explore neighborhood relationships, you usually want to include information up to a certain distance away from the center pixel. Using a square, or worse, a rectangular mask means that you are giving more weight to certain directions (where the 4 corners are pointing), which would only make sense for very specific images.