Sketch-based 3D Shape Retrieval Via Attention

How to retrieval 3D shapes effectively is a longstanding problem in the field of computer vision. The last years, with the popularity of touch devices such as mobile phones and tablets, sketch-based 3D shape retrieval has become a more efficient way of 3D shape retrieval. This paper proposes a sketch-based 3D shape retrieval framework via attention. First, each 3D shape is rendered into multiple images, and the important information of each view is enhanced through the attention network, while the other information of each view can be suppressed. This method is used as the basis for view fusion instead of choosing the best view. It helps to get comprehensive information about the 3D shape from multiple views, and can generate more robust 3D shape feature descriptors. Then the 3D shape features and sketch features are similarly learned to further narrow the gap between samples of the same categories in different domains, and increase the discrimination of samples of different categories in the same domain, which can better improve retrieval performance. The method in this paper has been verified on the dataset SHREC2013, confirming that the retrieval effect is better than some methods.