Part Detection for 3D Shapes via Multi-view Rendering

This paper presents a novel way of shape structure analysis namely part detection, which provides the positions and categories of all potential parts for any input shape. Different from segmentation, part detection does not provide precise shape part boundaries, but rather gives the bounding boxes and semantic labels of shape parts. This is useful for applications like style discovery, retrieval, visualization, etc. Part detection is achieved by multi-view rendering in this paper. Firstly, our method renders 3D shape into images under different perspectives, and detects the potential parts (as boxes) in the images using Faster R-CNN. Then, each detection result votes for the visible vertices in its own box under the corresponding perspective. Finally, we select the vertices for each part category according to the voting result, and generate the bounding boxes. The performance of this approach on a database of 16 shape categories is demonstrated in the end of paper.