Source-linked AI summary
Deep Hough Voting for Robust Global Registration
Junha Lee, Seungwook Kim, Minsu Cho, Jaesik Park
TL;DR
Pairwise point-cloud registration must align partially overlapping fragments, while existing learning-based pipelines have limitations in overlap assumptions, scalability, or iterative processing. The paper uses sparse 6D Hough voting with learnable refinement to estimate transformations from correspondence consensus. It reports state-of-the-art results on indoor benchmarks and multi-way ICL-NUIM, comparable KITTI performance, and faster speed than other learning-based pipelines.
Problem
Recent learning-based registration pipelines may require high overlap, lack scalability to large-scale point clouds, or rely on iterative post-processing.
Source: §1 · p. 1Method
The method casts candidate-correspondence votes in a sparse discretized 6D transformation space, refines the Hough space with a learnable module, and predicts the transformation from the maximum-vote consensus.
Source: §1 · p. 1Results
The method achieves state-of-the-art on 3DMatch and 3DLoMatch, performs comparably on KITTI, and sets a new state of the art on ICL-NUIM in a multi-way registration pipeline.
Source: §1 · p. 2Takeaways & Limitations
The approach is faster than other learning-based registration pipelines.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Point cloud registration is the task of estimating the rigid transformation that aligns a pair of point cloud fragments. We present an efficient and robust framework for pairwise registration of real-world 3D scans, leveraging Hough voting in the 6D transformation parameter space. First, deep geometric features are extracted from a point cloud pair to compute putative correspondences. We then construct a set of triplets of correspondences to cast votes on the 6D Hough space, representing the transformation parameters in sparse tensors. Next, a fully convolutional refinement module is applied to refine the noisy votes. Finally, we identify the consensus among the correspondences from the Hough space, which we use to predict our final transformation parameters. Our method outperforms state-of-the-art methods on 3DMatch and 3DLoMatch benchmarks while achieving comparable performance on KITTI odometry dataset. We further demonstrate the generalizability of our approach by setting a new state-of-the-art on ICL-NUIM dataset, where we integrate our module into a multi-way registration pipeline.
1. Introduction
Pairwise point-cloud registration must recover a rigid alignment despite partial overlap, while prior learning-based pipelines face overlap, scalability, or iterative-processing constraints. The paper proposes sparse 6D Hough voting with learnable refinement and reports strong benchmark results.
1. Introduction
- Registration estimates a rigid transformation aligning partially overlapping point-cloud fragments, while prior learning pipelines may assume high overlap, lack large-scale scalability, or require iterative post-processing.The task is described as important in autonomous driving, robotics, and augmented reality.Source: §1 · p. 1
- Sparse voting in a discretized 6D transformation space aggregates candidate-correspondence votes, then learnable refinement filters noise before the maximum-vote bin predicts the transformation.The sparse representation supports small bins for higher accuracy with less memory overhead than a dense parameter space.Source: §1 · pp. 1–2
- The approach sets new state of the art on real-world indoor 3DMatch [57]/3DLoMatch [30] and multi-way registration evaluated on, performs comparably on outdoor KITTI, and is faster than other learning-based registration pipelines.On 3DMatch, it is the most accurate overall and fastest among learning-based methods, and is about twice as fast as DGR [10].Source: Figure 1 · p. 1, §1 · p. 2, Figure 2 · p. 2
2. Related Work
Related work spans learned local features, robust correspondence filtering, Hough voting, and global or end-to-end registration. The proposed pipeline combines learned features with sparse Hough voting and single-shot refinement, avoiding iterative processing and intensive outlier rejection.
Learning-based local feature
Robust model fitting
- Robust fitting addresses outlier-heavy feature correspondences; RANSAC and variants sample candidate matches iteratively and evaluate resulting alignments..Source: §2 · p. 2
- Hough voting discretizes parameter space into bins, with prior uses spanning pose estimation, matching [8] [39], object detection [41], and image-category matching [38].The proposed approach applies sparse 6D voting to estimate transformations from correspondence consensus.Source: §2 · p. 2
Global registration frameworks
- Global registration alternatives include ICP, branch-and-bound, semi-definite programming, maximal clique selection [54], FGR, and SparseICP, with iterative or computational costs noted for several methods.FGR is faster but has comparable performance and sensitivity to hand-crafted feature configurations.Source: §2 · pp. 2–3
End-to-end registration frameworks
- The proposed pipeline combines a learning-based feature extractor, robust sparse Hough voting, and single-shot refinement, improving efficacy on large-scale real-world scans without iterative processing or intensive outlier rejection.This contrasts with prior end-to-end methods that may assume high overlap, scale poorly, or require iterative post-processing [52] [10].Source: §2 · p. 3
3. Method
The method turns local feature correspondences into triplet votes in a sparse 6D Hough space, refines noisy votes, and selects the maximum-vote bin to predict the transformation.
3. Method
- The network has three stages: local feature extraction, sparse Hough voting, and Hough-space refinement.These stages form the proposed registration pipeline.Source: §3 · p. 3
3.1. Local Feature Extraction
- FCGF features produce nearest-neighbor candidate correspondences, and a distance-based tuple test filters triplets with obviously spurious matches.The test uses a 3 × voxel-size distance threshold, relying on rotation and translation preserving point-to-point distances.Source: §3.1 · p. 3
3.2. Sparse Hough Voting
- Each triplet estimates a rotation and translation, then votes in a sparse 6D space whose bins encode the transformation parameters.Storing only bins that receive votes avoids redundant storage and accommodates unbounded translations; smaller bins support more accurate predictions.Source: §3.2 · p. 4
3.3. Learning to Refine Hough Space
- A U-shaped 6D sparse convolutional network refines the raw Hough space to address noise from outlier triplets and discretization.Training treats the bin containing the ground-truth transformation as the positive class.Source: §3.3 · pp. 4–5
3.4. Comparison with RANSAC
- Unlike RANSAC’s repeated hypothesis verification, the method accumulates votes in parallel and evaluates the consensus once after voting.The accumulated 6D sparse tensor also lets the refinement network use geometric patterns in transformation space.Source: §3.4 · p. 5
4. Experiments
The experiments evaluate pairwise registration on indoor and outdoor scans, then test the method in a multi-way pipeline. The method outperforms DGR on 3DMatch, performs strongly at low overlap and on KITTI, and generalizes to ICL-NUIM.
4.1. Datasets
- Experiments use 3DMatch indoor pairs with overlap as low as 30%, KITTI LiDAR scans, and Augmented ICL-NUIM with simulated depth noise.The 3DMatch split follows the provided train/test division; KITTI pairs are formed following the cited protocol.Source: §4.1 · p. 5
4.2. Implementation
- The experiments use pretrained FCGF features, train the Hough-space refinement network for ten epochs, and fix the feature extractor rather than training it end to end.CUDA parallelizes voting; Faiss supports feature matching, and MinkowskiEngine supports sparse tensors and convolutions.Source: §4.2 · p. 5
4.3. Evaluation metric
- Registration recall counts successful pairs, defined by RRE and RTE falling below preset thresholds; average RRE and RTE are computed on successful pairs.The authors also report these error averages regardless of registration success on 3DMatch because success-conditioned averages do not show overall prediction accuracy.Source: §4.3 · p. 6
4.4. Pairwise Registration
- On 3DMatch, the method outperforms DGR across all reported metrics at about twice its speed; on 3DLoMatch, it leads with both tested feature sets.Figure 4 shows recall across varying rotation and translation thresholds; Table 1 reports 3DMatch comparisons, while Table 3 covers 3DLoMatch.Source: Figure 4 · p. 6, §4.4 · pp. 6–7, Table 1 · p. 6
4.5. Outdoor LiDAR Registration
- On KITTI, the method surpasses the prior state of the art in recall and RRE and is over twice as fast as DGR, while FCGF + RANSAC has the lowest RTE.The proposed method runs seven times faster than FCGF + RANSAC.Source: §4.5 · p. 7, Table 4 · p. 7
4.6. Time comparison
- Registration times are approximately two times shorter on average for the proposed method, despite re-evaluated DGR times differing from those originally reported.DGR's reported and re-evaluated average times differ in opposite directions on 3DMatch and KITTI.Source: §4.6 · p. 7
4.7. Ablation Study
- Both Gaussian smoothing and Hough-space refinement improve results at both tested voxel sizes, but their combination does not consistently help; refinement yields larger gains.The method therefore uses the learnable refinement module alone; 2.5 cm voxels consistently outperform 5 cm voxels, although the method remains state of the art at 5 cm.Source: §4.7 · p. 8, Table 5 · p. 8
4.8. Multiway Registration
- In multi-way registration, the method outperforms state-of-the-art methods on all Augmented ICL-NUIM scenes except Living room 1, where it ranks second to DGR.The evaluation uses networks trained on 3DMatch and tested on the multi-way registration datasets.Source: §4.8 · p. 8, Table 6 · p. 8
5. Conclusion
The conclusion presents a point-cloud registration network combining dense feature extraction, sparse Hough voting, and learnable refinement. It reports state-of-the-art indoor results and comparable outdoor results, and identifies sampling improvements as future directions.
5. Conclusion
- The pipeline extracts dense features, uses sparse Hough voting, and applies learnable refinement to predict transformation parameters from correspondence consensus.The sparse parameter space is intended to provide high accuracy with low memory overhead.Source: §5 · p. 8
- The method is reported as highly robust to outliers, with state-of-the-art indoor results and comparable performance on the outdoor benchmark.The conclusion says this is achieved without intensive outlier filtering methods.Source: §5 · p. 8
- Keypoint selection and smarter triplet sampling are proposed as future directions for improving efficiency and performance.Source: §5 · p. 8
A. Appendix
The appendix provides additional details and results for the method.
A. Appendix
- The appendix collects supplementary methodological details and results.Source: A. Appendix · p. 12
A.1. Refinement module architecture
The refinement network processes sparse, noisy 6D Hough votes and outputs a refined Hough space. Its shared architecture uses a U-shaped sparse convolutional design with skip connections.
A.1. Refinement module architecture
- The refinement network is a 6D U-shaped sparse convolutional network that maps a sparse, noisy Hough space to a refined one.The same architecture is used across all experiments.Source: §A.1 · p. 12
- The network has seven convolutional layers, each followed by batch normalization and ReLU except the final layer.Source: §A.1 · p. 12
- Strided convolutions and transposed convolutions reduce and increase sparse-tensor cardinality by a factor of two, respectively.Source: §A.1 · p. 12
- Skip connections concatenate features from corresponding convolutional and transposed-convolutional layers along the channel dimension.The figure's example first convolution has kernel size 3, stride 1, and 32 output channels.Source: §A.1 · p. 12
A.2. Comparison with feature extractors
The appendix compares the method with feature descriptors and detectors under evaluation settings used in prior work, and specifies alternative successful-registration thresholds for 3DMatch and KITTI.
A.2. Comparison with feature extractors
- The appendix compares the method with recent 3D feature descriptors and detectors under evaluation settings used in [4].Source: §A.2 · p. 12
- For 3DMatch, the alternative success criterion is RMSE < 0.2 m, instead of the main manuscript's RTE < 0.3 m and RRE < 15°.The alternative criterion is specified for the appendix comparison.Source: §A.2 · p. 12
- For KITTI, the alternative thresholds are RTE < 2 m and RRE < 5°, while the main experiment uses RTE < 0.6 m.The table caption confirms the alternative KITTI thresholds.Source: §A.2 · p. 12
- The table's Size denotes the spatial hashing size in centimeters.Source: §A.2 · p. 12
A.3. Combination with FPFH feature
The method is evaluated with FPFH descriptors on 3DMatch and outperforms PointDSC with comparable efficiency.
A.3. Combination with FPFH feature
- The FPFH evaluation uses descriptors without finetuning because FCGF results are nearly saturated on 3DMatch.The reported timing excludes feature extraction and matching.Source: §A.3 · p. 12
- With FPFH descriptors on 3DMatch, the method outperforms PointDSC [3] with comparable efficiency.Source: §A.3 · p. 12
- Table a.2 reports the 3DMatch FPFH evaluation, with timing measured excluding feature extraction and matching.Source: §A.3 · p. 12
A.4. Additional scene-wise statistics
The additional analyses report scene-wise 3DMatch results, memory usage, and qualitative comparisons with DGR on 3DMatch and KITTI.
A.4. Additional scene-wise statistics
- The 3DMatch scene-wise statistics are reported per point-cloud pair and compared visually with baseline methods.Source: §A.4 · p. 12
- The method is consistently better than baselines across all unseen 3DMatch test scenes in recall, RE, and TE.Higher recall is better, while lower RE and TE are better; missing bars indicate zero successful alignments.Source: §A.4 · pp. 12–13
A.5. Analysis on memory usage
- The learnable refinement module uses slightly more memory than DGR [10], while the authors report a significant registration-accuracy boost.The comparison concerns peak memory usage and running time during evaluation on 3DMatch.Source: §A.5 · p. 13
A.6. Additional qualitative results
- Qualitative comparisons on 3DMatch and KITTI show accurate alignments on some point-cloud pairs that DGR [10] fails to align accurately.Figures a.3 and a.4 compare DGR and the proposed method side by side.Source: §A.6 · pp. 13–15