Leveraging Inlier Correspondences Proportion for Point Cloud Registration (2201.12094v2)

Abstract: In feature-learning based point cloud registration, the correct correspondence construction is vital for the subsequent transformation estimation. However, it is still a challenge to extract discriminative features from point cloud, especially when the input is partial and composed by indistinguishable surfaces (planes, smooth surfaces, etc.). As a result, the proportion of inlier correspondences that precisely match points between two unaligned point clouds is beyond satisfaction. Motivated by this, we devise several techniques to promote feature-learning based point cloud registration performance by leveraging inlier correspondences proportion: a pyramid hierarchy decoder to characterize point features in multiple scales, a consistent voting strategy to maintain consistent correspondences and a geometry guided encoding module to take geometric characteristics into consideration. Based on the above techniques, We build our Geometry-guided Consistent Network (GCNet), and challenge GCNet by indoor, outdoor and object-centric synthetic datasets. Comprehensive experiments demonstrate that GCNet outperforms the state-of-the-art methods and the techniques used in GCNet is model-agnostic, which could be easily migrated to other feature-based deep learning or traditional registration methods, and dramatically improve the performance. The code is available at https://github.com/zhulf0804/NgeNet.

• The paper introduces GCNet, a novel approach that improves point cloud registration by enhancing the proportion of inlier correspondences.
• It employs a pyramid hierarchy decoder and consistent voting strategy to extract and verify robust multi-scale features.
• Experimental results on 3DMatch and 3DLoMatch demonstrate registration recalls of 92.9% and 71.9%, underscoring its superior performance.

Overview of Leveraging Inlier Correspondences Proportion for Point Cloud Registration

The paper presents a novel approach to enhance the accuracy of point cloud registration by focusing on increasing the proportion of inlier correspondences. In point cloud registration, the challenge lies in determining a transformation that effectively aligns two overlapping fragments, a process crucial for applications such as 3D reconstruction and SLAM. The authors address the prevailing challenge of extracting discriminative features when dealing with partial and ambiguous surface data. This is achieved through the introduction of several key techniques which culminate in the development of the Geometry-guided Consistent Network (GCNet).

Key Contributions

The paper introduces three strategically designed techniques to boost the performance of feature-learning based point cloud registration:

1. Pyramid Hierarchy Decoder: This component characterizes point features across multiple scales, effectively facilitating the comprehensive capture of local and global structures within the point clouds.
2. Consistent Voting Strategy: By maintaining consistency in correspondences through cross-scale correspondence verification, this strategy is instrumental in pruning out outliers and enhancing the reliability of feature matching.
3. Geometry-Guided Encoding Module: This module incorporates geometric characteristics into the encoding process, enabling more precise feature extraction and subsequently increasing the likelihood of accurate point matching.

These techniques form the backbone of GCNet, which has been rigorously tested across diverse datasets, including indoor, outdoor, and object-centric synthetic ones. The results demonstrate a marked improvement over state-of-the-art methods, notably enhancing inlier correspondence ratios and registration recall.

Experimental Evaluation

The effectiveness of GCNet is substantiated through a comprehensive suite of experiments. The network outperforms existing methods on widely recognized datasets such as 3DMatch and its challenging variant 3DLoMatch, achieving registration recalls of 92.9% and 71.9%, respectively. Such results underscore the model's robustness and precision, particularly in low-overlap scenarios where many other methods falter.

Implications and Future Prospects

The implications of this work are pivotal for both practical deployments and theoretical exploration of point cloud registration. The model-agnostic nature of the proposed techniques means that they can be seamlessly integrated into other feature-based deep learning or traditional registration frameworks, promising substantial enhancements in alignment accuracy and computational efficiency.

Looking ahead, the paper suggests potential avenues for integrating the consistent voting mechanism and geometry-guided strategies into end-to-end deep learning architectures for further gains in registration performance. Additionally, investigating the scalability of these techniques for more complex, unstructured environments poses an interesting challenge, offering a rich vein of exploration for future research. This work lays a solid foundation for advancing the state-of-the-art in point cloud registration, paving the way for more robust applications in autonomous navigation, augmented reality, and beyond.