GeoTransformer: Fast and Robust Point Cloud Registration with Geometric Transformer (2308.03768v1)

Abstract: We study the problem of extracting accurate correspondences for point cloud registration. Recent keypoint-free methods have shown great potential through bypassing the detection of repeatable keypoints which is difficult to do especially in low-overlap scenarios. They seek correspondences over downsampled superpoints, which are then propagated to dense points. Superpoints are matched based on whether their neighboring patches overlap. Such sparse and loose matching requires contextual features capturing the geometric structure of the point clouds. We propose Geometric Transformer, or GeoTransformer for short, to learn geometric feature for robust superpoint matching. It encodes pair-wise distances and triplet-wise angles, making it invariant to rigid transformation and robust in low-overlap cases. The simplistic design attains surprisingly high matching accuracy such that no RANSAC is required in the estimation of alignment transformation, leading to $100$ times acceleration. Extensive experiments on rich benchmarks encompassing indoor, outdoor, synthetic, multiway and non-rigid demonstrate the efficacy of GeoTransformer. Notably, our method improves the inlier ratio by $18{\sim}31$ percentage points and the registration recall by over $7$ points on the challenging 3DLoMatch benchmark. Our code and models are available at \url{https://github.com/qinzheng93/GeoTransformer}.

• The paper introduces GeoTransformer, which enhances point cloud registration by encoding geometric features like pair-wise distances and triplet-wise angles.
• It employs a novel superpoint matching strategy using geometric self-attention and optimal transport for precise, RANSAC-free alignment.
• Numerical results demonstrate significant improvements in inlier ratios and registration recall, with up to 100x faster performance on benchmarks.

Overview of GeoTransformer: Fast and Robust Point Cloud Registration with Geometric Transformer

The paper presents a novel approach to point cloud registration, an essential task in computer graphics and robotics, through the introduction of the Geometric Transformer (GeoTransformer). This model enhances correspondence extraction for 3D point clouds by leveraging the transformer architecture enriched with geometric features.

Methodology

GeoTransformer stands out by avoiding reliance on keypoint detection, which can be challenging in low-overlap conditions. Instead, the model finds correspondences using downsampled superpoints, matching them based on neighboring patch overlaps. The key innovation is the adaptation of the Transformer to encode geometric features—specifically, pair-wise distances and triplet-wise angles—enabling transformation invariance and robustness in scenarios with low overlap.

The methodology of GeoTransformer includes:

1. Superpoint Sampling and Feature Extraction: Utilizes KPConv-FPN to downsample input clouds and extract features at multiple resolution levels, focusing on superpoints for initial matching.
2. Superpoint Matching Module: This module employs geometric self-attention and feature-based cross-attention to create hybrid features for reliable matching based on geometric consistency.
3. Geometric Self-Attention: Encodes intra-point-cloud geometric structures through transformation-invariant features derived from distances and angles, improving superpoint matching.
4. Point Matching Module: Propagates superpoint matches to dense points via optimal transport, enhancing the precision of point cloud correspondences.
5. RANSAC-Free Local-to-Global Registration: A method to estimate the alignment transformation efficiently, without depending on conventional RANSAC, thereby achieving substantial computational speedups.

Numerical Results and Implications

The model was extensively tested across various benchmarks, including indoor (3DMatch and 3DLoMatch) and outdoor (KITTI) datasets, as well as synthetic data (ModelNet40). GeoTransformer consistently achieved superior inlier ratios, feature matching recall, and registration recall, especially excelling in challenging low-overlap environments.

Key numerical results include:

• An improvement in inlier ratio by 18-31 percentage points on the challenging 3DLoMatch benchmark.
• A significant registration recall increase over state-of-the-art methods, maintaining accuracy with up to 100 times faster registration due to its RANSAC-free approach.

Theoretical and Practical Implications

The theoretical contribution of this work lies in its novel utilization of geometric feature encoding within transformers, bridging a gap in the transformation-invariant learning for point cloud registration. Practically, GeoTransformer’s acceleration of the registration process and its robustness in diverse scenarios may influence various applications, such as real-time 3D modeling and autonomous navigation.

Future Directions

The paper hints at directions for future research, including enhancing memory efficiency and adapting the model for non-rigid point cloud registration. Furthermore, exploring cross-modality registration and deep integration with semantic scene understanding could be transformative.

In summary, GeoTransformer offers a significant advancement in point cloud registration by merging geometric insight with the flexibility and power of transformer networks, providing both performance and efficiency improvements across a variety of challenging tasks and datasets.