Continuous conditional random field convolution for point cloud segmentation

Point cloud segmentation is the foundation of 3D environmental perception for modern intelligent systems. To solve this problem and image segmentation, conditional random fields (CRFs) are usually formulated as discrete models in label space to encourage consistency, which actually a kind postprocessing. In paper, we reconsider CRF feature point because it can capture structure features well improve representation ability rather than simply smoothing. Therefore, first model with continuous quadratic energy formulate its solution process message-passing graph convolution, by be easily integrated into deep network. We theoretically demonstrate that message passing convolution equivalent mean-field approximation model. Furthermore, build an encoder-decoder network based on proposed (CRFConv), CRFConv embedded decoding layers restore details high-level were lost encoding stage enhance location network, thereby benefiting segmentation. Analogous CRFConv, show classical also work collaboratively via another further results. Experiments various benchmarks effectiveness robustness method. Compared state-of-the-art methods, method achieve competitive performance.