Cognitive-inspired class-statistic matching with triple-constrain for camera free 3D object retrieval.

3D object retrieval has attracted much research attention in recent years, while most of the state-of-the-art approaches highly depend on the camera array settings for capturing 3D object views. We also note that the generality among objects in the same category has not been exploited in existing works. To this end, we propose a cognitive-inspired class-statistics matching method with triple-constraint (CSTC) for camera free 3D object retrieval. In this method, each object in the gallery set is represented by a free set of views without camera constraint. Inspired by representativeness heuristic, the category-independent distribution of each feature is calculated and Gaussian probabilistic models are generated with corresponding weights. Meanwhile, the distances between positive-to-positive examples are statistically measured based on pre-chosen matched views, and then the pairwise matching model is constructed in an off-line manner. In the retrieval procedure, for each query, the view-based distance measure is firstly converted into the object-based distance measure, and then the trained class-statistics matching models are employed to calculate the similarity between different objects, meanwhile, the constrains of the pairwise matching model are combined by CSTC model which can balances the performance and retrieval speed. In this model, since the object-based distance measure is firstly utilized, which is very helpful to speed up the retrieval, and then class-statistics matching model between the query object and gallery object, which can explore the generality among objects, is employed to improve the performance, moreover, the pairwise matching model is further used to filter the retrieval results, finally, in order to boost the retrieval speed and combine their complementary characteristic, their results are fused only by simple and effective linear combination (supervising good). We have conducted experiments on ETH, NTU-60, MVRED and PSB 3D datasets, and experimental results show that our performance outperforms or is comparable with the-state-of-the-art algorithms, but our retrieval speed obviously outperforms others.