CPFormer: contrast planes transformer for point cloud semantic segmentation

Point cloud multiplane mechanisms effectively capture rich geometric information from different projections, but this approach lacks explicit consistency constraints across planes, and the loose coupling mechanism of concatenation and summation leads to inconsistent cross-plane features. To address this challenge, we propose a contrast plane transformer, an innovative multiview network that utilizes contrastive learning to achieve cross-view feature coupling. Specifically, we first decompose the point cloud window into three orthogonal planes (XY, XZ, and YZ) and use the attention and graph-based local enhancement module to extract features from each plane. Subsequently, to realize cross-plane feature coupling, we propose a patch-level plane alignment mechanism. Specifically, we partition the planes into a certain number of patches, designate one patch from one plane (XY) as an anchor, and designate the corresponding patches from the other two planes as positive samples while utilizing other randomly selected patches as negative samples to construct an efficient and robust comparison mechanism. We not only achieve cross-plane consistency but also significantly reduce the computational complexity of point-level comparison. Finally, considering the imbalance of geometric information provided by different planes, we propose a dynamic plane fusion mechanism that can automatically adjust the weights of different planes according to the local geometric features of the input point cloud, thus realizing adaptive information fusion. Our method achieves 73% and 75.5% mIoU on the S3DIS and ScanNetv2 datasets, respectively, significantly outperforming most algorithms, which fully demonstrates the effectiveness of our algorithm.