Detail-preserving shape completion of point cloud models with articulated structure

Abstract

This paper advocates a novel deep-learning-based method for point cloud completion of multi-categorical articulated objects sharing the same topology. One popular approach for point cloud completion is to rely on a generic encoder-decoder architecture, where the feature maps of input are extracted with the critical set, which essentially consists of a set of points that play critical roles in the max-pooled features. But this pipeline has difficulties in retaining the local details, especially for arbitrary deformable, articulated objects of various categories, bringing category confused completion. In this paper, we propose a detail-preserving point cloud completion method for the complex articulated models by extracting features guided by their articulation topology with a fixed-order scheme, so as to accommodate both fine-grained categorical appearance and non-rigid deformation. First, we construct key subsets, which preserve both local, category-aware and global, non-rigid deformation features simultaneously for input sharing similar point densities, guided by a set of regressed key points approximating articulations. Second, we organize the key subsets with a fixed-order scheme during feature extraction to combat the possible interference due to diverse data component permutations during feature extraction, while upholding the algorithmic efficiency. Finally, we confirm in our evaluations that the new method completes general articulated point clouds with detailed categorical characteristics in high quality. We also show that after training on synthetic data, our method can be applied to real scan or web downloaded point clouds with similar point densities. Meanwhile, we built an Quadruped Point Cloud Completion (QPCC) dataset upon which new research topics could be further explored in geometry modeling and computer graphics.