Algorithm 3 describes a solution to find optimal weights using
FLDA. The algorithm is presented with a database of object and
clutter class training images with their corresponding labels. FLDA
outputs a linear projection vector whose slope corresponds to the
optimal weights assigned to the rectangles of the Haar-like feature.
Once the optimal weights are found, the optimal values for threshold
(ˆ
) and polarity terms (pˆ) are found by searching exhaustively
over all possible solutions.