[Paper Review] A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration

Renlang Huang, Yunxia Tang+2 more|arXiv (Cornell University)|Oct 14, 2024

3D Surveying and Cultural Heritage4 citations

TL;DR

This paper proposes CAST, a consistency-aware spot-guided Transformer for 3D point cloud registration that improves coarse matching accuracy and efficiency by enforcing geometric consistency through spot-guided and consistency-aware attention mechanisms. The method achieves state-of-the-art performance on both LiDAR and RGBD datasets with faster inference and robustness to low inlier ratios, eliminating reliance on optimal transport and RANSAC through a lightweight sparse-to-dense matching pipeline.

ABSTRACT

Deep learning-based feature matching has shown great superiority for point cloud registration in the absence of pose priors. Although coarse-to-fine matching approaches are prevalent, the coarse matching of existing methods is typically sparse and loose without consideration of geometric consistency, which makes the subsequent fine matching rely on ineffective optimal transport and hypothesis-and-selection methods for consistency. Therefore, these methods are neither efficient nor scalable for real-time applications such as odometry in robotics. To address these issues, we design a consistency-aware spot-guided Transformer (CAST), which incorporates a spot-guided cross-attention module to avoid interfering with irrelevant areas, and a consistency-aware self-attention module to enhance matching capabilities with geometrically consistent correspondences. Furthermore, a lightweight fine matching module for both sparse keypoints and dense features can estimate the transformation accurately. Extensive experiments on both outdoor LiDAR point cloud datasets and indoor RGBD point cloud datasets demonstrate that our method achieves state-of-the-art accuracy, efficiency, and robustness.

Motivation & Objective

• To address the limitations of existing coarse-to-fine point cloud registration methods that produce sparse, inconsistent coarse matches due to global cross-attention attending to irrelevant regions.
• To improve robustness and efficiency by enforcing geometric consistency in coarse matching through a novel consistency-aware matching confidence criterion.
• To eliminate dependency on computationally heavy methods like optimal transport and RANSAC by designing a lightweight, scalable fine matching module.
• To enable real-time applicability in robotics, such as odometry, through efficient sparse-to-dense registration and compatibility graph-based outlier rejection.
• To achieve state-of-the-art accuracy, efficiency, and robustness across diverse 3D point cloud datasets, including outdoor LiDAR and indoor RGBD scans.

Proposed method

• Introduces a spot-guided cross-attention module that restricts cross-attention to only relevant, confident neighboring patches based on a consistency-aware matching confidence criterion combining feature similarity and geometric compatibility.
• Proposes a consistency-aware self-attention module that enhances feature distinctiveness by aggregating information from salient nodes in a geometric compatibility graph.
• Employs multi-scale feature fusion to improve representation learning across different levels of geometric detail.
• Designs a lightweight fine matching module that supports both sparse keypoints and dense features, enabling efficient point-to-patch matching without optimal transport.
• Utilizes an efficient compatibility graph embedding module for robust outlier rejection, replacing traditional RANSAC-based hypothesis-and-selection pipelines.
• Integrates a sparse-to-dense registration pipeline that leverages the efficiency of sparse matching and the accuracy of dense refinement.

Experimental results

Research questions

• RQ1 Can a consistency-aware attention mechanism improve the reliability and geometric consistency of coarse point cloud correspondences?
• RQ2 How does spot-guided cross-attention reduce interference from irrelevant regions during feature matching?
• RQ3 To what extent can a consistency-aware self-attention module enhance feature distinctiveness through compatibility graph-based sampling?
• RQ4 Can a lightweight, sparse-to-dense matching module achieve high accuracy without relying on optimal transport or RANSAC?
• RQ5 Does the proposed method outperform existing state-of-the-art approaches in accuracy, efficiency, and robustness across diverse 3D point cloud datasets?

Key findings

• The proposed CAST model achieves state-of-the-art performance on both outdoor LiDAR and indoor RGBD point cloud registration benchmarks.
• CAST significantly improves coarse matching quality by reducing inconsistent correspondences through geometric consistency enforcement.
• The method achieves faster inference and better scalability than existing approaches relying on optimal transport or RANSAC.
• The consistency-aware self-attention and spot-guided cross-attention modules jointly enhance feature representation quality and matching reliability.
• The lightweight fine matching module enables accurate registration without the computational burden of patch-to-patch dense matching.
• The compatibility graph embedding module effectively rejects outliers, replacing inefficient hypothesis-and-selection pipelines with a more scalable solution.