Enhancing Performance of Point Cloud Completion Networks with Consistency Loss

TL;DR

This work tackles the one-to-many mapping problem in point cloud completion by introducing a completion consistency loss that enforces identical completion solutions across multiple incomplete views derived from the same object. The method comprises self-guided and target-guided consistency terms and a complete loss that combines these with the conventional reconstruction objective, optionally augmented by a density-aware Chamfer term, and can be plugged into existing PCCNs without network modifications. Empirically, the consistency loss yields consistent, significant improvements across architectures (e.g., PCN, AxFormNet, AdaPoinTr) and datasets (ShapeNet55/34, MVP), reduces generalization gaps to unseen categories, and even enables simpler, faster networks to rival more complex models while preserving inference speed. The approach demonstrates practical impact by delivering faster, more accurate point cloud completion and better robustness to noise and unseen shapes, with code and results to be released publicly.

Abstract

Point cloud completion networks are conventionally trained to minimize the disparities between the completed point cloud and the ground-truth counterpart. However, an incomplete object-level point cloud can have multiple valid completion solutions when it is examined in isolation. This one-to-many mapping issue can cause contradictory supervision signals to the network because the loss function may produce different values for identical input-output pairs of the network. In many cases, this issue could adversely affect the network optimization process. In this work, we propose to enhance the conventional learning objective using a novel completion consistency loss to mitigate the one-to-many mapping problem. Specifically, the proposed consistency loss ensure that a point cloud completion network generates a coherent completion solution for incomplete objects originating from the same source point cloud. Experimental results across multiple well-established datasets and benchmarks demonstrated the proposed completion consistency loss have excellent capability to enhance the completion performance of various existing networks without any modification to the design of the networks. The proposed consistency loss enhances the performance of the point completion network without affecting the inference speed, thereby increasing the accuracy of point cloud completion. Notably, a state-of-the-art point completion network trained with the proposed consistency loss can achieve state-of-the-art accuracy on the challenging new MVP dataset. The code and result of experiment various point completion models using proposed consistency loss will be available at: https://github.com/kaist-avelab/ConsistencyLoss .

Figures (7)

• Figure 1: Contradictory supervision signals could appear when an incomplete point cloud have multiple possible completion solutions, and could lead the network to fall into suboptimal solution regions. Point clouds are represented with solid lines in the figure for clarity.
• Figure 2: Two different incomplete point clouds that are obtained from one object should have the same solutions. Point completion network try to predict unseen point clouds on green dashed circle area.
• Figure 3: Completion results on the Shapenet55 dataset (test split).
• Figure 4: Completion results on the Shapenet34 dataset (test split - unseen).
• Figure 5: Comparison of performance for point completion networks AdaPointTr trained with Consistency Loss and without Consistency Loss on challenging dataset MVP Point Completion dataset test split.