Deep Confidence Guided Distance for 3D Partial Shape Registration

Dvir Ginzburg, Dan RavivAAAI 2022 · DOI

Abstract

Partial-to-partial 3D shape registration remains a significant challenge due to missing data, outliers, and internal symmetries. We present CGD-net, a deep learning framework that introduces a Confidence Guided Distance (CGD) metric. By fusing spatial proximity with learned feature similarity and employing a repulsion loss to distinguish symmetric parts, CGD-net achieves state-of-the-art performance on benchmarks like ModelNet40 and FAUST, effectively handling partial overlaps and complex geometries.

Introduction

3D registration—the task of finding the optimal rigid transformation to align two point clouds—is fundamental to computer vision and robotics. While classic methods like Iterative Closest Point (ICP) are effective for full-shape alignment, they often fail in "partial-to-partial" scenarios where point clouds only partially overlap. Two primary hurdles persist: the presence of non-overlapping outliers and the "internal symmetry" problem, where local features of distant parts (e.g., left vs. right wings of an airplane) appear identical, leading to incorrect alignments.

Methodology

CGD-net addresses these challenges through a three-pillar approach:

  1. Confidence Guided Distance (CGD): Unlike the standard Chamfer Distance, CGD incorporates a "confidence score" derived from learned feature embeddings. This ensures that points are only considered a match if they are both spatially close and feature-consistent.
  2. Hierarchical Feature Extraction with Repulsion Loss: The network utilizes an EdgeConv-based architecture for local geometry capture. A novel repulsion loss is introduced to force spatially distant segments to have distinct embeddings, breaking the ambiguity caused by symmetry.
  3. Confidence-Based Sampling: Instead of random sampling, the model learns to identify points with high matching probability. It then evaluates multiple alignment hypotheses in parallel to find the global optimum.
The CGD-net architecture featuring hierarchical feature extraction and confidence-guided sampling.

Experiments and Results

The model was evaluated on ModelNet40 (synthetic CAD) and FAUST (real-world human scans), modified for ~40% overlap. CGD-net demonstrates a significant performance leap over existing methods like RPM-net and PRNet. On ModelNet40, it achieved a rotation RMSE of 3.12, compared to 6.01 for PRNet. On the highly articulated FAUST dataset, the improvement was even more pronounced (3.81 vs 6.32 RMSE).

Qualitative comparison of registration results on human and airplane models. CGD-net successfully aligns symmetric parts where other methods fail.

The results highlight that symmetry confusion is primarily a feature descriptiveness problem. By forcing the network to encode global context into local features via repulsion loss, CGD-net effectively distinguishes between identical-looking symmetric parts.

Discussion and Conclusion

While highly effective, CGD-net's current reliance on a dense alignment matrix poses scalability challenges for very large point clouds (e.g., >100k points). Future work will explore hierarchical soft alignment schemes to reduce computational complexity while maintaining registration accuracy. In summary, CGD-net provides a robust solution for partial 3D registration by integrating confidence-guided metrics and symmetry-aware feature learning.

Summary updated

BibTeX
@inproceedings{ginzburg2022deep,
  author = {Dvir Ginzburg and Dan Raviv},
  title = {{Deep Confidence Guided Distance for 3D Partial Shape Registration}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i1.19951},
}
Figure