[Paper Review] CoFiNet: Reliable Coarse-to-fine Correspondences for Robust Point Cloud Registration

Hao Yu, Lin Fu+3 more|arXiv (Cornell University)|Oct 26, 2021

3D Shape Modeling and Analysis49 references95 citations

TL;DR

CoFiNet proposes a detection-free, coarse-to-fine framework that learns hierarchical correspondences for robust 3D point cloud registration, outperforming state-of-the-art with fewer parameters, especially in low-overlap scenarios.

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.

Motivation & Objective

• Motivate robust point cloud registration without relying on repeatable keypoint detection.
• Introduce a coarse-to-fine correspondence framework to shrink the search space for refinement.
• Develop a density-aware, differentiable refinement module to produce point-level correspondences.
• Show that coarse-scale node matching guided by overlap improves recall and robustness.
• Demonstrate superior performance with fewer parameters on indoor and outdoor benchmarks.

Proposed method

• Down-sample dense point clouds to uniformly distributed nodes with shared KPConv encoders and learn joint node features.
• Use a Correspondence Proposal Block with self-, cross-, and self-attention to produce a coarse similarity matrix between down-sampled nodes and apply Sinkhorn optimization with slack entries to obtain coarse node correspondences.
• Decode strengthened features to dense point descriptors for refinement.
• Expand coarse node proposals to patches via point-to-node grouping and apply a density-adaptive matching module to refine correspondences by solving a differentiable optimal transport problem within patches.
• Compute a joint coarse-and-fine loss: L = Lc + λLf with specialized supervision for coarse overlap-aware weighting and fine density-aware refinement.

Experimental results

Research questions

• RQ1 Can a coarse-to-fine, detection-free pipeline improve repeatability and robustness of 3D point correspondences for registration?
• RQ2 Does a overlap-guided coarse matching reduce the search space and improve efficiency without sacrificing accuracy?
• RQ3 Can a density-aware refinement module provide reliable point-level correspondences across varying point densities?
• RQ4 How does CoFiNet perform against state-of-the-art methods on indoor (3DMatch, 3DLoMatch) and outdoor (KITTI) benchmarks, especially under low overlap?
• RQ5 What are the parametric efficiency and ablation insights when using fewer but higher-quality correspondences?

Key findings

• CoFiNet achieves superior Registration Recall and Feature Matching Recall across indoor and outdoor benchmarks, notably on 3DLoMatch where overlap is limited.
• The method uses significantly fewer parameters than competing approaches while delivering strong performance.
• CoFiNet yields high Inlier Ratios even with fewer correspondences, indicating well-calibrated confidence scoring.
• The density-adaptive refinement module robustly handles varying point densities within patches, improving point-level correspondence accuracy.
• A coarse-overlap weighting mechanism effectively shrinks the refinement search space and enhances robustness in low-overlap scenarios.