Low-overlap point cloud registration algorithm based on coupled iteration

We present BC-PCNet, a P oint C loud registration model based on B idirectional C oupled iteration. The proposed model addresses the challenge of registering point clouds with low overlap. We introduce a new supervisory signal called “Mask Region.” This signal is used to supervise the overlapping region of the two point clouds during the coupled iterative process, enhancing the accuracy of registration. We also improve the registration accuracy by increasing the number of coupled iterative steps. Moreover, by randomly downsampling the non-overlapping part of the point cloud, we reduce the amount of input training data and increase the speed of model training and registration. Compared to the latest models, our model performs well for low-overlap point cloud registration. Experiments show that BC-PCNet achieves a 0.6%/4.1% improvement in recall precision on the 3DMatch/3DLoMatch datasets.