Geometric Transformer for Fast and Robust Point Cloud Registration (2202.06688v2)

Abstract: We study the problem of extracting accurate correspondences for point cloud registration. Recent keypoint-free methods bypass the detection of repeatable keypoints which is difficult in low-overlap scenarios, showing great potential in registration. 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 to learn geometric feature for robust superpoint matching. It encodes pair-wise distances and triplet-wise angles, making it robust in low-overlap cases and invariant to rigid transformation. 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. Our method improves the inlier ratio by $17{\sim}30$ percentage points and the registration recall by over $7$ points on the challenging 3DLoMatch benchmark. Our code and models are available at https://github.com/qinzheng93/GeoTransformer.

Geometric Transformer for Fast and Robust Point Cloud Registration

Point cloud registration is an essential task in 3D graphics, computer vision, and robotics, involving the alignment of two partially overlapping point clouds using rigid transformations. This paper presents a novel approach, termed the Geometric Transformer (GeoTransformer), specifically designed to enhance the accuracy and efficiency of point cloud registration.

Methodology

The paper tackles key challenges in point cloud registration by bypassing traditional keypoint detection, which is difficult in low-overlap scenarios, and instead focuses on a keypoint-free method. The approach relies on matching downsampled superpoints and propagating these matches to dense points. Unlike other models, this method does not employ RANSAC, enabling significant acceleration in computation.

The GeoTransformer utilizes a geometric self-attention mechanism to encode structural information and capture transformation-invariant features. The innovation lies in the use of pair-wise distances and angles within point triplets, allowing the model to understand context without being affected by geometric transformations.

Key Components

1. Superpoint Sampling and Feature Extraction: The method begins by extracting features at multiple resolution levels using KPConv-FPN, where certain layers correspond to superpoints.
2. Geometric Self-Attention: This component integrates global geometric context and captures the inherent structure of point clouds through distance and angular embeddings. It effectively replaces positional embeddings that are transformation-variant.
3. Point Matching: Using an optimal transport layer for dense point correspondences, this stage ensures robust mapping and alignment between the matched superpoints.
4. Local-to-Global Registration (LGR): The transformation is estimated via a parameter-free approach that foregoes RANSAC, significantly reducing computation time while maintaining accuracy.

Experimental Results

The paper presents comprehensive evaluations on the 3DMatch and KITTI benchmarks, demonstrating superior performance compared to several state-of-the-art methods. The results exhibit consistent improvements in matching recall and registration accuracy, particularly in low-overlap cases — a scenario where traditional methods often struggle. The Geometric Transformer showed registration recall improvement of over 7 percentage points on the challenging 3DLoMatch dataset and provided up to 30 percentage points increase in inlier ratios.

Implications and Future Work

The introduction of a geometry-aware Transformer architecture reveals significant practical implications, particularly in applications requiring fast and reliable point cloud registration without the computational overhead of RANSAC. The transformation-invariant feature extraction opens avenues for robust registration solutions across various scales and conditions.

The paper suggests future exploration in cross-modality registration, potentially extending the method's capability to harmonize 2D and 3D data sources. The integration with semantic scene understanding could also be explored to further enhance the robustness and contextual awareness of the registration process.

Conclusion

This work presents a substantial advancement in the field of point cloud registration by addressing the inherent limitations of existing methods through a geometry-centric approach. The proposed GeoTransformer not only accelerates the registration process but also increases robustness and accuracy, particularly in scenarios characterized by minimal overlap. The insights and methodologies introduced in this paper hold significant potential for future developments in 3D data processing and intelligence systems.