Traditionally, the rectangles of a Haar-like feature with two rectangles
are assigned default weights of 1 and −1. This yields a projection
line that is parallel to the vector 1 − 2. To train a weak
classifier as in (3), a scalar value is found such that object and
clutter points are best separated. As shown in Fig. 5(b), divides
the SRFS into two regions; the object region, coded white, and the
clutter region, coded black.
Fi