Learning feature manifold from user's relevance feedback for 3D model retrieval

High-dimensional feature vectors extracted from 3D models always lie on a manifold embedded in the original Euclidean space. A novel relevance feedback method using geodesic distance to discover the intrinsic manifold structure was proposed. Meanwhile, a model potential theory for revised SVM was proposed to improve the relevance feedback mechanism. Experimental results show that the approach is effective in improving the performance of content-based model retrieval systems.