I found the answer to the problem: opencv "cv.floodFill" function does exactly what I want, and is muche faster than scipy watershed (by a factor of ~100).