Fig. 15.

Comparison of baseline (top) and fin-space identification scheme (bottom). The two paradigms are illustrated proceeding from left to right. By associating descriptor matching scores (left column) to reference locations in a global fin space (colouration), the improved scheme (bottom) accumulates information not into a single, class-specific scalar (top approach), but forms a scoring vector that encodes the pattern of matchings over fin space. Identity is then judged via a random forest based on the learned reliability of the matching patterns