PaperlayerView exact revision on arXiv

Source-linked AI summary

Register Anything Model for Generalizable and Robust Point Cloud Registration

Zheng Qin, Junhua Xi, Yan Huang, Jiahong Lai, Anwen Huang, Qiong Li, Guangda Zhang, Kai Xu

arXiv:2610.28821v1cs.CV

TL;DR

Registration methods are often tailored to one domain, while differences in scale, density, noise, and structure complicate generalization across domains. RAM trains a unified model using geometry-aware normalization and a hierarchical registration pipeline with an efficient geometric transformer. Across 12 benchmarks, it reports state-of-the-art performance and outperforms BUFFER-X on novel-domain Registration Recall.

  • Problem

    Learning-based registration methods are commonly tailored to a specific domain, and substantial geometric differences make generalization to novel domains challenging.

    Source: §1 · pp. 1–2

  • Method

    RAM trains a unified cross-domain model that adaptively voxelizes and normalizes point clouds, then uses hierarchical correspondence refinement and a limited-scope geometric transformer.

    Source: §1 · p. 2, §3.2 · p. 5, §3.3 · p. 6

  • Results

    RAM outperforms BUFFER-X in Registration Recall on novel domains and achieves an average improvement of 10.2 percentage points in Inlier Ratio.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Geometry-aware normalization eliminates the need to manually select a domain-specific voxel size at inference time.

    Source: §4.5 · p. 9

    RAM’s KPConv-based backbone remains rotation-variant and may struggle under extreme scenarios.

    Source: §5 · p. 10

Abstract

from arXiv · show

We present Register Anything Model (RAM), a unified model for generalizable 3D point cloud registration that robustly estimates 6-DoF transformations across diverse domains spanning sensors and environments. Existing registration methods are typically tailored to a single domain, with carefully tuned architectures and hyperparameters, and thus degrade markedly when applied to novel domains. We bridge this gap by training a single model on diverse data from indoor, outdoor, and object domains. A central challenge is the severe geometric discrepancy across domains, i.e., differences in scene scale, sensor noise, point density, and structural patterns, which hinders processing with a unified point-based model. To address this, we introduce a geometry-aware normalization strategy that jointly voxelizes and normalizes point clouds into a standardized geometric space, adaptively based on their geometric properties to preserve salient local structures while standardizing resolution. Built on this normalization, RAM follows a hierarchical coarse-to-fine paradigm, and employs an efficient geometric transformer that models local geometric consistency instead of globally over all points. This design mitigates discretization noise introduced by voxelization, and improves computational efficiency. Extensive experiments on 12 benchmarks demonstrate state-of-the-art performance in terms of inlier ratio and registration recall. Notably, our method outperforms the counterparts trained on each separate domain, showing superior generalization capability.

1 Introduction

Registration methods trained for a single domain often struggle on novel domains because point-cloud geometry varies substantially. RAM trains one model across domains, combining geometry-aware normalization and an efficient geometric transformer; experiments report strong results across 12 benchmarks.

1 Introduction

  • Single-domain learning-based methods often struggle on novel domains because sensor and environmental differences create substantial shifts in geometric characteristics.These shifts are associated with inaccurate correspondences and degraded registration accuracy.Source: §1 · pp. 1–2
  • RAM jointly voxelizes and normalizes point clouds into a standardized geometric space, choosing the voxel size adaptively to preserve salient local structures.The strategy is framed as identifying a large voxel size that does not materially alter geometry, then normalizing to a reference resolution.Source: §1 · p. 2
  • RAM uses a limited-scope geometric transformer to model inter-point-cloud consistency, mitigating voxelization noise while reducing computational cost.The design seeks to balance feature discriminativeness and computational efficiency across diverse conditions.Source: §1 · pp. 2–3
  • RAM follows a hierarchical coarse-to-fine registration paradigm, first extracting superpoint correspondences and then refining them into point correspondences.The model is trained jointly on data from indoor, outdoor, and object domains.Source: §1 · pp. 2–3
  • RAM outperforms BUFFER-X in Registration Recall on novel domains and achieves an average 10.2 percentage-point improvement in Inlier Ratio.The experiments evaluate RAM on 12 benchmarks after training on four registration datasets spanning indoor, outdoor, and object domains.Source: §1 · pp. 2–3

2 Related Work

Prior registration work includes matching-based pipelines, hierarchical correspondence refinement, and transformation estimation, while generalizable methods still face cross-domain geometry shifts. Cross-source registration concerns point clouds from different acquisition sources, whereas cross-domain registration concerns pairs spanning environments or sensors; RAM’s pipeline normalizes geometry before correspondence extraction and transformation recovery.

2.1 Point Cloud Registration

  • Hierarchical registration methods first find coarse superpoint or patch correspondences and refine them into fine-grained point correspondences.Other learning-based approaches use keypoints, rotation-robust operators, robust estimators, or direct transformation regression.Source: §2.1 · p. 3

2.2 Generalizable Point Cloud Registration

  • Domain-specific geometric shifts limit generalization, and adaptation methods such as BUFFER-X still have inaccurate correspondences in novel domains and require domain-specific tuning.BUFFER-X adapts voxel size using sphericity and search radius using point density when generalizing a model trained on 3DMatch.Source: §2.2 · p. 3

2.3 Cross-source Point Cloud Registration

  • Cross-source registration aligns clouds from different acquisition sources, while cross-domain registration develops a unified model for pairs spanning environments or sensors.RAM’s pipeline adaptively voxelizes and normalizes clouds before hierarchical feature extraction, superpoint matching, dense correspondence refinement, and transformation recovery.Source: §2.3 · p. 3, Figure 2 · p. 4

3 Method

RAM estimates rigid transformations through a coarse-to-fine registration pipeline. It standardizes point-cloud geometry and uses local geometric consistency to support processing across domains.

3.1 Overview

  • The pipeline extracts correspondences between normalized point clouds and estimates the rotation and translation that align them.Hierarchical point features come from a KPConv-FPN backbone, with coarsest-level points serving as superpoints and half-resolution points supporting dense matching.Source: §3.1 · p. 4, §3.3 · p. 5

3.2 Geometry-aware Normalization

  • Geometry-aware normalization selects voxel sizes from point-cloud geometry and rescales clouds to consistent scales and densities for a unified search radius.The method iteratively enlarges voxel size while geometric error remains below tolerance, using linearity and planarity changes to assess distortion.Source: §3.2 · p. 5
  • Voxel size begins from the smaller cloud’s bounding-box diagonal, and local covariance eigenvalues characterize linearity, planarity, and sphericity during normalization.Geometric error uses the maximum relative linearity and planarity error per seed and the median over the half of seeds with the largest sphericity.Source: §3.2 · p. 5

3.3 Register Anything Model

  • RAM first matches superpoints, then refines them into dense point correspondences before estimating the rigid transformation.Dense matching uses hybrid point features, overlapping scores, and feature similarity; mutual top-k selection extracts local correspondences.Source: §3.3 · pp. 5–6
  • EGeoTR replaces global geometric consistency modeling with local self- and cross-attention, reducing embedding memory complexity to O(Nkd) and avoiding domain-specific downsampling ratios.The global embedding has O(N^2d) memory complexity, and distant-superpoint attention scores are empirically low.Source: §3.3 · pp. 5–6

3.4 Loss Functions

  • RAM combines overlap-aware circle loss for superpoint features, negative log-likelihood loss for confidence matrices, and binary cross-entropy loss for overlap scores.The overall objective is a weighted sum of the three losses.Source: §3.4 · p. 6

4 Experiments

RAM is evaluated across indoor, outdoor, and object registration benchmarks, with a hybrid-trained model tested against domain-specific and generalization baselines. Results and ablations show strong performance across domains, while visualizations illustrate robustness to challenging geometry.

4.1 Benchmarks and metrics.

  • The evaluation covers 12 indoor, outdoor, and object benchmarks; RAM is trained on a hybrid set, and results use Inlier Ratio, error, and Registration Recall metrics.The hybrid training set combines 3DMatch, KITTI, nuScenes, and ModelNet.Source: §4.1 · pp. 6–7

4.2 Experiments on Indoor Benchmarks

  • 3.4 pp and 5.4 pp higher Registration Recall than BUFFER-X on ScanNet++ and TIERS, respectively; TIERS Inlier Ratio is also 5.2 pp higher.RAM consistently outperforms both BUFFER-X variants on these unseen indoor benchmarks.Source: §4.2 · p. 7, Table 1 · p. 7

4.3 Experiments on Outdoor Benchmarks

  • RAM achieves the best performance on all metrics across four unseen outdoor benchmarks, while outdoor-only training outperforms the hybrid model without nuScenes on most.The unseen benchmarks are ETH, WOD, MIT, and Oxford.Source: §4.3 · p. 8, Table 2 · p. 8

4.4 Experiments on Object Benchmarks

  • RAM consistently outperforms the object-registration baselines on ModelNet and ShapeNet, although GeoTrans achieves better Inlier Ratio.BUFFER-X variants fail to achieve reasonable results on these object benchmarks.Source: §4.4 · p. 8, Table 3 · p. 9

4.5 Ablation Studies

  • The hybrid model performs best across all ablation benchmarks; overlap scores improve matching, the transformer matters, and geometry-aware normalization usually beats static voxel sizes.Geometry-aware normalization also removes the need to choose a domain-specific voxel size at inference.Source: §4.5 · p. 9, Table 4 · p. 9

4.6 Qualitative Results

  • Visualizations show RAM extracting accurate correspondences across domains and remaining robust to incomplete geometry, duplicated objects, and symmetric structures.The comparisons include BUFFER and BUFFER-X.Source: §4.6 · p. 10, Figure 3 · p. 10

5 Conclusion

The paper presents RAM as a unified model for cross-domain point cloud registration, with experiments supporting generalizability across indoor, outdoor, and object scenarios. The authors also identify constraints involving rotation, scale, and distant outliers.

5 Conclusion

  • Experiments across diverse indoor, outdoor, and object scenarios demonstrate strong cross-domain generalizability and robustness, with state-of-the-art performance.The conclusion describes RAM as a unified registration model.Source: §5 · p. 10
  • The KPConv-based backbone is rotation-variant and may struggle under extreme rotations; RAM may also not scale to extremely large point clouds.The authors suggest tiling large scenes and processing tiles separately, potentially with overlap and streaming.Source: §5 · p. 10
  • Distant outliers can inflate the initial voxel size and oversmooth geometry; a separate outlier-removal step could mitigate this.Source: §5 · p. 10

A.1 Implementation Details

Implementation uses a reference voxel size and a scale-dependent initialization, alongside a KPConv-FPN backbone and a hybrid-data training procedure.

A.1 Implementation Details

  • The method sets reference voxel size vref = 0.025 m and defines ϕ as a monotonically increasing piecewise function with anchors ϕ(1)=0.015 m, ϕ(5)=0.025 m, ϕ(50)=0.1 m, and ϕ(200)=0.3 m.Source: §A.1 · p. 13
  • The backbone has four KPConv-FPN stages, halves point-cloud resolution at each stage, and doubles feature channels.KPConv search radius is 2.5 times the voxel size at each stage; EGeoTR uses 128 neighbors per superpoint.Source: §A.1 · pp. 13–14
  • RAM is trained for 3M iterations on hybrid data from 3DMatch, KITTI, nuScenes, and ModelNet, without balancing dataset samples.Geometry-aware normalization is omitted during training for efficiency; extra random scaling ranges from 0.8 to 1.2.Source: §A.1 · p. 14

A.2 Benchmarks

The benchmarks span indoor reconstruction, LiDAR-based outdoor scenes, and CAD objects, with test sets assembled using dataset-specific pair selection and splits.

A.2 Benchmarks

  • 3DMatch uses 62 RGB-D scenes and standard train, validation, and test splits; 3DLoMatch pairs have 10–30% overlap, compared with over 30% for 3DMatch test pairs.The 3DMatch test sets contain 1,623 pairs and the 3DLoMatch test set 1,781 pairs.Source: §A.2 · p. 14
  • ScanNet++ testing uses 2,050 iPhone 13 Pro depth-image point-cloud pairs, while TIERS testing uses 870 pairs from several LiDAR scanners.The TIERS data come from Velodyne VLP-16, Ouster OS1-64, and Ouster OS0-128 scans.Source: §A.2 · p. 14
  • KITTI and nuScenes use 64-beam and 32-beam LiDAR, respectively; their test sets contain 555 and 1,330 pairs.KITTI pairs are selected from sequences 8–10, while nuScenes pairs are generated using the KITTI procedure.Source: §A.2 · p. 14
  • Outdoor tests include 713 ETH pairs, 130 WOD pairs, 230 sparse MIT pairs, and 301 Oxford pairs.MIT pairs are drawn from an outdoor campus sequence captured with a Velodyne VLP-16.Source: §A.2 · pp. 14–15
  • ModelNet testing uses 1,146 CAD models from held-out categories, while ShapeNet testing samples 50 models from each of 55 categories.ShapeNet point clouds are uniformly sampled to 8,192 points per model.Source: §A.2 · p. 15

A.3 Metrics

The section defines benchmark-specific Registration Recall criteria and uses stricter thresholds than BUFFER-X.

A.3 Metrics

  • For 3DMatch and 3DLoMatch, successful registration means registration error below 0.2m.Source: §A.3 · p. 15
  • For other benchmarks, success requires both relative rotation error below τrre and relative translation error below τrte.Source: §A.3 · p. 15
  • The evaluation uses stricter thresholds than BUFFER-X, with benchmark-specific values listed in Tab. 5.Source: §A.3 · p. 15

A.4 Additaional Results

Additional tests report RAM's inference efficiency, performance under extended training and large rotations, cross-source generalization, and statistical comparisons on ScanNet++ and Oxford.

Computational efficiency

  • RAM has the fastest model-inference time among the compared methods, and pipelined stages yield around 4 fps.The current normalization implementation rebuilds raw-point KD-trees at every iteration; caching or pre-voxelization could reduce this overhead.Source: §A.4 · pp. 15–16

Volume of training data

  • The hybrid model improves 3DMatch and 3DLoMatch results over the indoor-only model even when training exposures are approximately matched.The comparison extends indoor-only training to 120 epochs; the authors attribute the gains mainly to cross-domain training rather than sample count alone.Source: Table 8 · p. 16, §A.4 · p. 16

Rotation invariance

  • On 3DMatch and 3DLoMatch, large rotations reduce RAM's Inlier Ratio by 0.8 and 1.0 points and Registration Recall by 0.7 and 2.7 points, respectively.RAM still outperforms PARENet in Inlier Ratio on both benchmarks under large rotations.Source: Table 9 · p. 16, §A.4 · p. 16

Cross-source registration

  • RAM remains behind Cross-PCR on cross-source registration, partly because its current encoder is sensitive to severe density and sampling discrepancies.Source: §A.4 · p. 16
  • RAM substantially outperforms GeoTransformer on both 3DCSR subsets, especially the more challenging LiDAR subset, despite not being trained specifically for cross-source registration.Source: §A.4 · p. 16, Table 10 · p. 17

Statistical test

  • RAM achieves the lowest median RTE and RRE and the smallest variance on ScanNet++ and Oxford, with statistically significant improvements over all tested baselines.Source: Figure 4 · p. 17, §A.4 · p. 17

A.5 Discussions

The discussion finds that cross-domain training can improve broad and per-domain performance, while domain-relevant data and geometric diversity contribute complementary benefits.

Trade-off between generalization and per-domain accuracy

  • The hybrid model consistently outperforms corresponding domain-specific models across evaluated benchmarks under the same architecture and optimization protocol.The authors report no systematic trade-off between generalization and per-domain accuracy under this training design.Source: §A.5 · p. 17

Domain-relevant training data

  • The complete hybrid model performs best overall, while removing nuScenes particularly reduces MIT performance and an outdoor-only model partly recovers it.On MIT, the outdoor-only model raises Inlier Ratio from 62.5 to 72.3 and RR from 97.4 to 98.7.Source: §A.5 · p. 18

Complementary roles of training data

  • Cross-domain diversity and target-domain data play complementary roles: broader geometric patterns support generalization, while target data supply domain-specific structures and sampling characteristics.Source: §A.5 · p. 18

Effect of LiDAR sparsity on low-overlap registration

  • With approximately matched training exposures, the hybrid model improves 3DLoMatch Inlier Ratio from 50.3 to 53.7 and Registration Recall from 79.2 to 81.1.This supports the authors' view that geometric diversity, rather than sample exposure alone, primarily drives the improvement.Source: §A.5 · p. 18
  • LiDAR sparsity weakens local geometric cues, whereas low overlap reduces shared structure and increases correspondence ambiguity; most outdoor training pairs still have relatively high overlap.Source: §A.5 · p. 18
  • Sparse, incomplete LiDAR observations enrich training's geometric and sampling patterns and may indirectly improve robustness to partial observations and low-overlap registration.Source: §A.5 · p. 18

A.6 Broader Impacts

The broader-impacts discussion positions RAM as a cross-domain registration approach with potential uses in multiple 3D applications.

A.6 Broader Impacts

  • RAM is presented as enabling spatial alignment across diverse sensors without extensive per-domain fine-tuning or retraining.Source: §A.6 · p. 18
  • Potential applications include LiDAR odometry, autonomous navigation, AR/VR, digital twins, construction monitoring, and cultural heritage preservation.Source: §A.6 · p. 18
  • The authors suggest geometry-aware normalization could serve as a component toward point cloud foundation models.Source: §A.6 · p. 18

Related papers

Loading 2610.28821v1…