Graph-Guided Deformation for Point Cloud Completion

Jieqi Shi, Lingyun Xu, Liang Heng, Shaojie Shen

Abstract

Captured 3D point clouds from LiDAR or depth sensors are frequently incomplete due to occlusions and limited sensor resolution. Conventional deep learning approaches treat point cloud completion as a generative task, often losing fine geometric details present in the partial input. This paper proposes a novel framework that reframes completion as a graph-guided deformation problem. By utilizing the partial input as "control points" and employing a Graph Convolutional Network (GCN) to refine a coarse intermediate shape, the model preserves local intricacies while hallucinating missing regions. Our approach achieves state-of-the-art results on synthetic benchmarks and demonstrates significant robustness on sparse, real-world LiDAR datasets.

Introduction

In autonomous driving and robotics, 3D sensing is fundamental. However, raw point clouds are often sparse and incomplete due to sensor limitations and environmental occlusions. Existing methods typically employ an encoder-decoder architecture to compress partial inputs into a global shape code, from which a complete object is generated from scratch. This bottleneck often results in "average" shapes that lack the high-fidelity local details present in the original partial scan.

To address this, we propose treating the known input points as reliable anchors. Rather than generating an object entirely from scratch, we use these points to guide the deformation of a coarse template, ensuring that the final output respects the observed geometry while completing the missing parts.

Methodology

Our proposed pipeline consists of two primary stages designed to mimic traditional mesh deformation techniques in computer graphics.

1. Shape-Preserved Generation Network

The first stage utilizes a multi-resolution encoder and a tree-structured decoder to produce a coarse, intermediate point cloud. This provides a global structural estimate of the complete object, serving as a rough template for the subsequent refinement.

2. Graph-Guided Deformation Network

The core contribution of this work is the deformation stage. The original input points are designated as "control points", while the intermediate generated points serve as "supporting points". We construct a k-Nearest Neighbor (k-NN) graph to establish Laplacian coordinates, which represent local geometric relationships.

The proposed two-stage pipeline: a Generation Network produces a coarse shape, followed by a Graph-Guided Deformation Network that refines the geometry using input points as anchors.

A Graph Residual Network (G-ResNet) then optimizes point positions by simulating a least-squares Laplacian deformation. The training objective combines Chamfer Distance for global accuracy, a Matching Loss to ensure control points remain fixed, and a Shape-Preserving Loss to maintain local smoothness along the graph edges.

Experiments and Results

We evaluated our model on the synthetic ShapeNet dataset and real-world datasets including KITTI and a self-collected drone-mounted LiDAR dataset (Pandar40).

Quantitative Performance

On ShapeNet, our method achieved a per-point Chamfer Distance of $6.02 \times 10^{-4}$, significantly outperforming prior state-of-the-art models such as SA-Net ($7.74 \times 10^{-4}$), TopNet ($9.72 \times 10^{-4}$), and PCN ($14.72 \times 10^{-4}$). Ablation studies confirmed that the deformation stage is critical, as removing it increased the error to $7.56 \times 10^{-4}$.

Qualitative Analysis

As shown in the comparisons, our model excels at preserving fine-grained details that baseline models often blur into solid blobs. For instance, in car reconstructions, features like roofs and wheels are more distinct and faithful to the input data.

Qualitative comparison of completion results on the ShapeNet dataset. Our method preserves structural details more effectively than PCN and TopNet.

On real-world LiDAR data from KITTI and Pandar40, the model demonstrated strong zero-shot generalization, successfully reconstructing complex objects under various noise and sparsity conditions. Furthermore, using our completed point clouds as input for a PointNet classifier improved accuracy to 94.58%, compared to 90.00% for PCN.

Conclusion

By reframing point cloud completion as a deformation task, we effectively bridge the gap between global shape hallucination and local detail preservation. While the density of control points presents a trade-off between detail and global coverage, our graph-guided approach sets a new benchmark for high-fidelity 3D reconstruction. Future work will explore dynamic control point selection to further enhance robustness across varying levels of input sparsity.

Summary updated

BibTeX
@article{shi2021graph,
  author = {Jieqi Shi and Lingyun Xu and Liang Heng and Shaojie Shen},
  title = {{Graph-Guided Deformation for Point Cloud Completion}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2021},
  doi = {10.1109/lra.2021.3097081},
}
Figure