Improving Shape Retrieval by Fusing Generalized Mean First-Passage Time

摘要

In recent years, many efforts have been made to fuse different similarity measures for robust shape retrieval. In this paper, we firstly propose generalized mean first-passage time (GMFPT) that extends the mean first-passage time (MFPT) to the general form. Instead of focusing on the propagation of similarity information, GMFPT is introduced to improve pairwise shape distances, which denotes the mean time-steps for the transition from one state to a set of states. Through a semi-supervised learning framework, an iterative approach with a time-invariant state space is further proposed to fusing multiple distance measures, and the relative objects on the geodesic paths can be gradually and explicitly retrieved. The experimental results on different databases demonstrate that shape retrieval results can be effectively improved by the proposed method.