A Powerful Partial Relevance Feedback for 3D Model Retrieval

Relevance feedback(RF) has been proved to be an effective way to improve the precision and recall of 3D model retrieval. However, the existing RF approaches do not consider which local part of the feedback example is similar or dissimilar with the query model although they recorded whether the whole model is similar or not. In this study, a partial relevance feedback(PRF) method which overcomes this deficiency is discussed . First, an improved silhouette based descriptor is proposed to satisfy the PRF method. Second, a new mathematical model for partial relevance feedback is set up and optimal solution is also given: a SVM based classifier is trained to classify the models; the variables which have influence on similarity measurement are optimized to minimize the average distance between the query model and the feedback examples, and then the similarities between the query model and all the models in database are recalculated. At last, Some experiments are given to illustrate the outperformance of the proposed method over the other methods.