Abstract
Real-world 3D scans often suffer from incompleteness due to occlusions and sensor limitations. While supervised methods excel at point cloud completion, they require large-scale paired datasets that are difficult to obtain for real-world scenes. This paper introduces USSPA, an unsupervised framework that leverages inherent object symmetry and a novel refinement autoencoder to complete partial point clouds. By integrating a symmetry learning module with a shape-preserving architecture, USSPA achieves high-fidelity completion without paired ground truth. Experimental results on real-world benchmarks demonstrate that USSPA outperforms state-of-the-art unsupervised methods in both geometric accuracy and multi-category generalization.
Introduction
3D scanning technologies like LiDAR frequently produce partial point clouds, hindered by occlusions and noise. Traditional supervised completion models rely on synthetic paired data, which often fails to generalize to the complexities of real-world scans. Unsupervised methods offer a promising alternative but frequently struggle with shape distortion, loss of fine detail, and category-specific training requirements. USSPA addresses these challenges by utilizing symmetry as a geometric prior and maintaining the integrity of the original input points through a shape-preserving autoencoding process.
Methodology
The USSPA architecture is built upon a Generative Adversarial Network (GAN) framework designed for unsupervised learning, focusing on structural consistency and detail preservation.
Key components of the system include:
- Symmetry Learning Module: This core innovation automatically identifies an object's symmetry plane to mirror existing data. By leveraging the fact that most man-made objects are symmetric, the module fills large gaps based on structural priors rather than pure hallucination.
- Refinement Autoencoder: Following initial coarse prediction, this module employs Local Feature Grouping and an Upsampling Refinement Module to enhance point density and ensure surface smoothness.
- Dual Discriminators: A Point Discriminator ensures realistic spatial distribution, while a Feature Discriminator validates the latent representation against known complete object distributions.
- Classifier-Guided Training: Unlike models restricted to single categories, USSPA incorporates category labels into the discriminator's feedback, allowing it to learn features for diverse objects like chairs, tables, and sofas simultaneously.
Datasets and Evaluation
The model was trained using ShapeNet for complete object distributions and ScanNet for real-world partial inputs. To rigorously evaluate performance, the authors developed a new benchmark using Scan2CAD to align clean models with real scans, providing a reliable pseudo-ground truth for quantitative analysis. The model was also tested on the PCN Dataset to verify its generalization capabilities.
Results and Discussion
USSPA demonstrates significant improvements over existing unsupervised techniques in both quantitative and qualitative metrics. Quantitatively, it achieved a Chamfer Distance of 8.56 on real-scene benchmarks, significantly outperforming the previous state-of-the-art unsupervised method, "Unpaired," which scored 10.47.
Qualitatively, as shown in the comparison above, the model excels at preserving thin structures and sharp edges, such as chair legs and table corners, which are often blurred or lost by competing methods. Furthermore, USSPA shows robust performance in multi-category scenarios and provides useful unsupervised classification features, achieving 69.8% accuracy on real-world scans, surpassing PointNet++ (58.6%) on the same task.
Conclusion
USSPA presents a robust solution for unsupervised point cloud completion by effectively combining symmetry priors with local feature refinement. By explicitly fusing original input points with generated data, the model ensures that observed geometry is preserved while missing parts are reconstructed with high fidelity. While currently dependent on clean CAD models for distribution learning, USSPA sets a new standard for reconstruction in complex real-world 3D scenes. Future work will explore moving toward completion using purely raw, unpaired real-world data without CAD assistance.
Summary updated