Revision history [back]

Cascade classification will always return you a bounding box result, which matches to the windows that has found a model match. The convex detection actually looks for edge points like the fingertips and such, giving you the possibility to calculate finger orientations.

So a first step could be a cascade classifier to detect the initial hand, then lay over a covex hull or a snake like approach, then calculate features and orientations. Keeping track of the hand can be done by following the palm hand with a tracker mechanism.