A New Performance Benchmark for Content-Based 3D Model Retrieval

At first, the paper introduces the most prevailing 3D model benchmark, the Princeton shape benchmark. Deficiencies emerged in the benchmark are then discussed in depth, which are concluded as: 1) models belonging to the same category are not exactly similar according to their shapes, and 2) category similarity is totally ignored. To overcome those shortcomings, the paper proposes a new model classification method, based on which a novel retrieval performance metric, GSSS (get score from similarity sequence), is designed and discussed. Experimental results have shown that GSSS is better than the precision-recall benchmark on most occasions.