Boosting 3D model retrieval with class vocabularies and distance vector revision

Visual-based 3D model retrieval presents great potential, for its broad application prospects and relatively high accuracy. A branch of visual-based methods utilizes scale invariant feature transform (SIFT) on 2D rendered images of a 3D model viewed from regularly sampled locations on a sphere, and then the bag-of-words framework is employed to improve the retrieval precision. However, in existing research literature, the universal vocabulary is usually trained from all the considered classes of models in database, which ignores the significant class information. To overcome this problem, we present a novel 3D model retrieval algorithm based on class vocabularies (CV-3DMR), which uses the category information of the classified database. Concretely, the class vocabularies are obtained through the adaptation of the universal vocabulary using class-specific data and the maximum a posterior (MAP) criterion. To boost the retrieval accuracy, we propose a distance vector revision strategy based upon the primary query results in top ranking. This strategy could be popularized to other approaches directly to promote their retrieval performance. Experimental results on the Princeton Shape Benchmark show that the proposed method makes a significant improvement over the compared 3D model retrieval algorithms.