High-level semantic feature for 3D shape based on deep belief networks

Deep learning has emerged as a powerful technique to extract high-level features from low-level information, which shows that hierarchical representation can be easily achieved. However, applying deep learning into 3D shape is still a challenge. In this paper, we propose a novel high-level feature learning method for 3D shape retrieval based on deep learning. In this framework, the low-level 3D shape descriptors are first encoded into visual bag-of-words, and then highlevel shape features are generated via deep belief network, which facilitates a good semantic preserving ability for the tasks of shape classification and retrieval. Experiments on 3D shape recognition and retrieval demonstrate the superior performance of the proposed method in comparison to the state-of-the-art methods.