A structural-constraint 3D point clouds segmentation adversarial method

Abstract

Listen

Translate article

Point cloud segmentation is a key task of shape analysis with various applications. Existing segmentation methods usually apply a single segmentation network to compute point-wise loss for network training. The point-wise loss, which will lose some structure information and resulting in non-consistency labeling. To deal with this problem, we propose a segmentation adversarial framework. Compared with the existing segmentation methods, the proposed method extends the single segmentation network to a complex network by adding a discrimination network, which can compensate adversarial structural loss and enforce spatial label consistency. The discrimination network is added to the single segmentation network, without increasing any computational burden in testing process as it is not required during inference. Moreover, a structural-constraint point cloud segmentation adversarial network is proposed. Especially, in feature-encoding process, we capture the fix-scale neighborhood of each point and concern the relationship between point pairs to provide better local context. Additionally, the proposed method realizes the global constraint through condition setting and the consistency of prediction label by using learnable loss function. Experimental results show the proposed method can produce more reasonable and consistency segmentation and labeling results.