Figure 4

We sample from the exemplar databases to produce a distribution over the training data given some input descriptor. In this case, the descriptor is the probability of skin/nonskin pixels in the face image shown at the top left. The PCA decomposition of the descriptor is used to decide how to traverse the binary tree, branching depending on the sign once the median has been subtracted (to balance the tree). At each branching of the tree a randomness factor is computed (based on a Gaussian) which results in the leaf nodes of the tree being explored. The leaf nodes are indices into the database which, in turn, point to specific frames in a sequence. We show here the illustrative matches generated for 10 samples with associated probabilities.