[Paper Review] P2C: Self-Supervised Point Cloud Completion from Single Partial Clouds

Ruikai Cui, Shi Qiu+5 more|arXiv (Cornell University)|Jul 27, 2023

3D Shape Modeling and Analysis41 references4 citations

TL;DR

P2C introduces the first self-supervised point cloud completion framework that trains on single partial scans per object, eliminating the need for complete shapes or multiple views. By using region-aware Chamfer distance and normal consistency constraints, it achieves state-of-the-art performance on both synthetic and real-world data, outperforming methods trained with complete ground truth.

ABSTRACT

Point cloud completion aims to recover the complete shape based on a partial observation. Existing methods require either complete point clouds or multiple partial observations of the same object for learning. In contrast to previous approaches, we present Partial2Complete (P2C), the first self-supervised framework that completes point cloud objects using training samples consisting of only a single incomplete point cloud per object. Specifically, our framework groups incomplete point clouds into local patches as input and predicts masked patches by learning prior information from different partial objects. We also propose Region-Aware Chamfer Distance to regularize shape mismatch without limiting completion capability, and devise the Normal Consistency Constraint to incorporate a local planarity assumption, encouraging the recovered shape surface to be continuous and complete. In this way, P2C no longer needs multiple observations or complete point clouds as ground truth. Instead, structural cues are learned from a category-specific dataset to complete partial point clouds of objects. We demonstrate the effectiveness of our approach on both synthetic ShapeNet data and real-world ScanNet data, showing that P2C produces comparable results to methods trained with complete shapes, and outperforms methods learned with multiple partial observations. Code is available at https://github.com/CuiRuikai/Partial2Complete.

Motivation & Objective

• To address the challenge of point cloud completion when complete 3D shapes are unavailable or expensive to obtain.
• To eliminate reliance on paired partial-complete data, multiple views, or complete shape supervision in training.
• To develop a self-supervised framework that learns structural priors from single partial observations of the same object category.
• To improve completion quality by incorporating geometric priors such as local planarity and shape consistency.
• To unify weak supervision cues from single-view partial scans into a robust, scalable training paradigm.

Proposed method

• The method groups incomplete point clouds into local patches to model continuous surface regions, enabling masked patch prediction.
• It introduces Region-Aware Chamfer Distance (RCD), which computes distance only for points near observed regions, avoiding penalization of missing parts and constraining outliers.
• A latent reconstruction loss based on cycle consistency ensures that completing different partial regions of the same object yields consistent full shapes.
• The Normal Consistency Constraint (NCC) regularizes predictions by minimizing variance in normal vector similarity among nearby points, enforcing local surface continuity.
• The framework combines RCD and NCC with a standard reconstruction loss to guide completion while preserving shape completeness and reducing artifacts.
• The model is trained end-to-end using only single partial scans per object, with no need for complete ground truth or multi-view alignment.

Experimental results

Research questions

• RQ1 Can point cloud completion be effectively trained using only a single partial scan per object, without complete shape supervision?
• RQ2 How can a self-supervised framework learn structural priors from diverse missing parts across a category of objects?
• RQ3 What loss functions are most effective in preventing overfitting to observed regions while enabling recovery of missing geometry?
• RQ4 How can geometric priors like local planarity be incorporated to improve surface continuity in completion results?
• RQ5 Can a self-supervised method outperform supervised or weakly-supervised baselines when complete shapes are not available?

Key findings

• P2C achieves state-of-the-art performance on both synthetic ShapeNet and real-world ScanNet datasets, even surpassing models trained with complete ground truth.
• The ablation study shows that adding Region-Aware Chamfer Distance (RCD) reduces Chamfer Distance by 5.1 compared to the baseline, indicating improved completion capability.
• Incorporating the latent reconstruction loss further improves performance, reducing CD-ℓ₂ by 1.5 compared to RCD-only training.
• The Normal Consistency Constraint (NCC) leads to the best overall result, achieving a CD-ℓ₂ of 11.2 on the 3D-EPN dataset, outperforming alternatives that use mean or minimum normal similarity.
• The model significantly reduces outlier generation and improves surface continuity, as validated by qualitative and quantitative analysis on real-world scans.
• The framework demonstrates strong generalization, producing high-quality completions on real-world data despite being trained solely on single partial scans.