CoFiNet: Reliable Coarse-to-fine Correspondences for Robust Point Cloud Registration (2110.14076v1)

Abstract: We study the problem of extracting correspondences between a pair of point clouds for registration. For correspondence retrieval, existing works benefit from matching sparse keypoints detected from dense points but usually struggle to guarantee their repeatability. To address this issue, we present CoFiNet - Coarse-to-Fine Network which extracts hierarchical correspondences from coarse to fine without keypoint detection. On a coarse scale and guided by a weighting scheme, our model firstly learns to match down-sampled nodes whose vicinity points share more overlap, which significantly shrinks the search space of a consecutive stage. On a finer scale, node proposals are consecutively expanded to patches that consist of groups of points together with associated descriptors. Point correspondences are then refined from the overlap areas of corresponding patches, by a density-adaptive matching module capable to deal with varying point density. Extensive evaluation of CoFiNet on both indoor and outdoor standard benchmarks shows our superiority over existing methods. Especially on 3DLoMatch where point clouds share less overlap, CoFiNet significantly outperforms state-of-the-art approaches by at least 5% on Registration Recall, with at most two-third of their parameters.

Analysis of CoFiNet for Robust Point Cloud Registration

The research paper "CoFiNet: Reliable Coarse-to-fine Correspondences for Robust Point Cloud Registration" proposes a novel approach to the challenging task of establishing correspondences in point cloud registration. The authors introduce CoFiNet, a Coarse-to-Fine Network, which extracts correspondences between point clouds without relying on traditional keypoint detection methods. This work addresses the common issue of repeatability in keypoint detection, offering a more robust alternative for point cloud registration used in applications such as scene reconstruction and autonomous driving.

Methodological Details

CoFiNet employs a hierarchical architecture to extract correspondences from coarse to fine scales. At the coarse level, the model focuses on matching down-sampled nodes whose vicinity points share significant overlap. This approach effectively reduces the search space for subsequent finer scale operations. At the finer scale, the identified overlap areas are expanded into patches, where correspondences are refined by a density-adaptive matching module designed to handle varying point densities. This hierarchical approach is enabled by a novel weighting scheme, which prioritizes local overlap ratios to guide coarse-scale matching, and a differentiable optimal transport problem which performs refined correspondence matching.

To achieve this, CoFiNet leverages shared KPConv encoders for initial down-sampling and feature learning. The key innovative aspect is the aggregation of global context using an attention mechanism, allowing the network to strengthen learned features prior to matching. Furthermore, CoFiNet incorporates a slack variable to handle unmatched nodes effectively, enhancing its reliability in registering point clouds with low overlap, as evidenced by its performance on 3DLoMatch and KITTI benchmarks.

Numerical Results and Claims

The paper presents strong numerical results, particularly on the 3DLoMatch benchmark, where CoFiNet demonstrates its effectiveness by outperforming state-of-the-art methods by over 5% in Registration Recall, while utilizing only two-thirds of the parameters. These results highlight the efficiency of CoFiNet, which achieves this performance with a notably reduced model size compared to other contemporary approaches.

Interestingly, CoFiNet's performance is underscored by its high Inlier Ratio, which reflects the quality of the correspondences it extracts. Additionally, when registration is evaluated without the robustness provided by RANSAC, the CoFiNet's superior correspondent extraction capability is clearly evident, indicating its potential for direct applications in scenarios without significant post-processing.

Implications and Future Directions

The paper's findings have important implications for the domain of point cloud registration, particularly in challenging real-world scenarios where overlapping among data points is minimal. CoFiNet's ability to bypass issues related to repeatability in keypoint detection marks a substantial development in robust point cloud registration. This contribution is expected to benefit applications such as SLAM and 3D scene reconstruction, where reliable and efficient registration processes are essential.

Future developments could explore integrating CoFiNet with other neural network architectures to further enhance adaptability and performance across varying environments. Additionally, addressing limitations such as handling non-distinctive regions and improving the sparsity of extracted point correspondences would further bolster CoFiNet’s application scope. The introduction of explicit schemes to reject outliers at the coarse scale could also streamline its precision and applicability.

Overall, the CoFiNet presents a substantial advancement in the field of point cloud registration, offering a reliable, parameter-efficient, and adaptable method suitable for both indoor and outdoor applications, setting a strong foundation for further exploration and enhancement in 3D computer vision tasks.