Point cloud registration can be solved by searching for correspondence pairs. Searching for correspondence pairs in human body point clouds poses some challenges, including: (1) the similar geometrical shapes of the human body are difficult to distinguish. (2) The symmetry of the human body confuses the correspondence pairs searching. To resolve the above issues, this article proposes a Hierarchical Tolerance Mask Correspondence (HTMC) method to achieve better alignment by tolerating obfuscation. First, we define various levels of correspondence pairs and assign different similarity scores for each level. Second, HTMC designs a tolerance loss function to tolerate the obfuscation of correspondence pairs. Third, HTMC uses a differentiable mask to diminish the influence of non-overlapping regions and enhance the influence of overlapping regions. In conclusion, HTMC acknowledges the presence of similar local geometry in human body point clouds. On one hand, it avoids overfitting caused by forcibly distinguishing similar geometries, and on the other hand, it prevents genuine correspondence relationships from being masked by similar geometries. The codes are available at https://github.com/ChenPointCloud/HTMC.
Keywords: 3D point matching; Computer vision; Point cloud registration.
© 2023 Yu et al.