Distance Metric Learning Based on Semantic Correlation Strength for 3D Model Retrieval

Abstract

3D model retrieval is an important part of multimedia information retrieval. To overcome the drawbacks of traditional text-based method, current researches mainly concentrate on the content-based 3D model retrieval. However, the effect of the content-based method is not satisfactory because of the semantic gap. Therefore, we propose a new 3D model retrieval method using semantic-correlation-strength-based distance metric learning. The method firstly obtains semantic correlation strength between 3D models from users' long-term relevance feedbacks, then uses semantic correlation strength as weights and adopts improved weighted relevant component analysis method to learn a Mahalanobis distance function. Finally, using the learned Mahalanobis distance metric function to retrieve 3D models. Experiments on Princeton Shape Benchmark show the effectiveness of our proposed method.