Manifold Ranking for Sketch-Based 3D Model Retrieval

The demand for 3D model retrieval is increasing, and the sketch-based method has been proven to be the most effective and efficient approach to retrieve 3D models. The existing methods calculate distance based on feature extraction, showing its limitation in improving retrieval accuracy. Thus, a second ranking making use of relevance between features is a good way to go. In this paper, an extended manifold ranking method is presented as a new retrieval framework. Line drawings are abstracted to represent 3D models, and a visual vocabulary is used to describe the local features of both sketches and line drawings. To rank the similarities between models, a method of semantic classification as a constraint is presented. We use similarity weight to control the classification difference between models so that the ranking score of models that belong to the same class holds a higher similarity weight. Furthermore, based on the idea of manifold learning, a KNN algorithm is adopted to obtain better ranking results. Experiments on standard testing datasets have demonstrated that the proposed algorithm significantly improves the accuracy of 3D model retrieval and outperforms current state-of-the-art algorithms by comparison.