Evaluation of relevance feedback methods for 3D model retrieval

Relevance feedback as a powerful search engine technique bridges the gap between high-level semantic knowledge and low-level object representation. In this paper, we experimentally evaluate 5 state-of-the-art relevance feedback methods: Elad2001, Space Warping, Linear Discriminant Analysis (LDA), Biased Discriminant Analysis (BDA) and Support Vector Machine (SVM). In order to guarantee the experiments reproductive, they are assessed based on the best 3D model descriptor DESIRE and the publicly available 3D model database Princeton Shape Benchmark (PSB). The experiments show that the retrieval performance of 3D model search engine may be significantly improved with the application of relevance feedback. In contract to the ambiguous results comparing SVM and BDA from previous paper, SVM was found to outperform BDA with distinct advantage, and they were followed by Elad2001, LDA and Space Warping.