인천대학교 인디고클랫폼 인천대학교 기술지원허브

Pointcloudcompletionaddressestheproblemofreconstructing complete 3D geometric shapes from incomplete or partial observations. While deep learning-based methods have achieved success using Chamfer Distance (CD) as a loss function, inherent limitations remain. The nearest-neighbor correspon dence mechanism of CD allows many-to-one mappings between point clouds, leading to distributional imbalance and undesired clustering artifacts. This distributional bias accumulates during training without being captured by CD metrics, resulting in misalignment with ground-truth distributions. To mitigate this issue, we (i) propose categorization to distinguish aligned from misaligned point correspondences, and (ii) introduce a correspondence-aware loss function consisting of two complementary terms—an uncovered region guidance loss and an interior-spreading loss—that provide explicit relocation guidance for misaligned points. Consequently, our approach achieves improved completion and better alignment with ground truth distributions. The proposed loss reuses correspondence information already computed during CD, maintaining computational efficiency comparable to that of CD while avoiding the exponential overhead of EMD.Throughexperimentsconductedacrossfivepointcloudcompletionarchitecturesandthreebenchmark datasets, we demonstrate that the proposed loss function improves completion models quantitatively and qualitatively.