Register Anything Model for Generalizable and Robust Point Cloud Registration

AI summary

Overview

Research area: 3D computer vision and robotics — specifically, learning-based point cloud registration (estimating the rigid 6-DoF transformation that aligns one 3D point cloud to another).

Technical level: Advanced. The paper assumes familiarity with correspondence-based registration pipelines, point-cloud backbones (KPConv-FPN), geometric transformers, and standard benchmarks such as 3DMatch and KITTI. The plain-language sections below make the core ideas accessible without that background.

Scope: The paper proposes Register Anything Model (RAM), a single registration network trained on indoor, outdoor, and object data that generalizes to 12 benchmarks without domain-specific tuning.

What This Paper Is About

Most point cloud registration models are built and tuned for one domain — indoor rooms, outdoor LiDAR scans, or single objects — and degrade sharply when applied to a different one, because scene scale, point density, sensor noise, and structural patterns differ so much. The authors ask whether one model trained on data from many domains can register point clouds anywhere, and they answer yes, provided the input geometry is first standardized. The goal is a single set of weights that produces accurate correspondences and reliable 6-DoF transformations across sensors and environments without per-domain hyperparameter tuning.

Key Contributions

1. Cross-domain training study and a unified model. The authors show that jointly training on indoor (3DMatch), outdoor (KITTI, nuScenes), and object (ModelNet) data yields a single registration model that generalizes better than models trained on each domain separately, and they evaluate it on 12 benchmarks.

2. Geometry-aware normalization. A module that jointly voxelizes and normalizes each point cloud pair into a standardized geometric space, adaptively choosing a voxel size per pair by measuring how much voxelization changes local linearity and planarity, then rescaling to a reference voxel size. This removes the need to hand-pick a voxel size (for example 2.5 cm indoors versus 30 cm outdoors) at inference time.

3. Efficient geometric transformer (EGeoTR). A transformer that models inter-point-cloud geometric consistency only within a local neighborhood of k nearest superpoints, reducing the memory complexity of geometric structure embedding from O(N²d) to O(Nkd). This avoids the domain-specific downsampling that global geometric attention would otherwise require.

4. State-of-the-art results across 12 benchmarks. RAM reports the best or near-best inlier ratio and registration recall in indoor, outdoor, and object settings, including gains over BUFFER-X of 3.4 and 5.4 percentage points in registration recall on unseen ScanNet++ and TIERS, and 5.2 percentage points in inlier ratio on TIERS.

Main Findings

• Cross-domain training beats single-domain training. The hybrid-trained model outperforms the indoor-only, outdoor-only, and object-only variants on all seven benchmarks used in the ablation (ScanNet++, TIERS, ETH, WOD, MIT, Oxford, ShapeNet). The object-only model, for example, reaches 28.8% registration recall on TIERS versus 90.0% for the hybrid model.

• Unseen indoor benchmarks improve over BUFFER-X. On ScanNet++ and TIERS, RAM reaches 70.6% / 90.0% and 66.0% / 90.0% (inlier ratio / registration recall), versus 83.5% / 86.6% and 54.4% / 81.5% for the 3DMatch-pretrained BUFFER-X. RAM wins on recall on both and on inlier ratio on TIERS, while BUFFER-X retains a higher inlier ratio on ScanNet++; the authors attribute this to BUFFER's dedicated learnable outlier-rejection module.

• Strongest on the hardest indoor benchmark. On 3DLoMatch, RAM achieves 81.1% registration recall against 74.3% for GeoTrans and 74.3% for the 3DMatch-pretrained BUFFER-X, indicating robustness to severe occlusion and low overlap.

• Best across all six outdoor benchmarks. RAM reports 90.7% inlier ratio / 99.5% recall on KITTI, 70.5% / 97.4% on nuScenes, 97.5% / 100.0% on ETH, 91.2% / 100.0% on WOD, 74.2% / 100.0% on MIT, and 82.0% / 99.3% on Oxford. On KITTI this is a 6.6 percentage-point inlier-ratio gain over the baselines.

• Large margin on object benchmarks. On ModelNet, RAM reaches 65.6% inlier ratio / 94.2% recall with RRE 1.875° and RTE 0.023 m; on ShapeNet, 65.9% / 79.6% with RRE 4.592° and RTE 0.047 m. Both BUFFER-X variants collapse here — the 3DMatch-pretrained one reaches only 13.3% recall on ModelNet with RRE 95.483° — which the authors present as evidence of the difficulty of transferring scene registration to objects.

• Overlap scores and the transformer both matter for point matching. Removing overlap-score prediction lowers ScanNet++ recall from 90.0% to 88.9% and inlier ratio from 70.6% to 68.3%; additionally removing the transformer drops these to 63.7% and 88.9%.

• Adaptive normalization beats static voxel sizes. With identical weights, geometry-aware normalization beats the static-voxel counterpart on most benchmarks, with the largest gaps on unseen domains such as ETH (97.5% versus 62.4% inlier ratio) and TIERS (66.0% versus 47.8%). Oxford is an exception where the static setting scores higher on inlier ratio (86.0% versus 82.0%).

• Removing nuScenes hurts less than removing domain relevance. A variant trained without nuScenes still beats baselines on most benchmarks except MIT, but the outdoor-only model outperforms it on most metrics, showing that domain-relevant data still helps within a domain.

Methodology in Plain English

The pipeline has three stages.

1. Standardize the geometry. Outdoor LiDAR scans span roughly 100 m and are sparse; indoor scans span under 10 m and are dense. A fixed neighborhood radius cannot suit both. The authors instead decide, per point cloud pair, how coarsely they can voxelize before the shape actually changes. They start from a fraction of the bounding-box diagonal, voxelize, sample seed points, and compare local shape descriptors (linearity, planarity, sphericity, derived from covariance eigenvalues) before and after voxelization. If the median error over the most "spherical" — that is, geometrically complex — patches stays under a tolerance, they enlarge the voxel and repeat. The final voxel size defines a scale factor that rescales the cloud to a reference resolution. Every cloud then enters the network with comparable scale and density.

2. Match coarse, then refine. A KPConv-FPN backbone extracts hierarchical features. At the coarsest level, superpoints are matched using a transformer that attends only within each superpoint's k nearest neighbors and exchanges information across the two clouds, producing a correlation matrix from which the top matches are selected. At half resolution, dense points inside each matched superpoint pair are compared, with a lightweight version of the same transformer plus a predicted per-point overlap score that suppresses non-overlapping regions.

3. Estimate the transform. Dense correspondences are fed to a robust estimator (LGR) that solves for rotation and translation. Training uses three losses: an overlap-aware circle loss on superpoint features, a negative log-likelihood loss on the confidence matrices, and a binary cross-entropy loss on overlap scores.

Why This Matters

Impact on research. The paper reframes cross-domain registration from "adapt a pretrained model to a new domain" (the BUFFER / BUFFER-X line of work) to "train once on many domains." It also shows that a normalization step can substitute for the domain-specific architectural choices — voxel sizes, downsampling ratios, search radii, RANSAC thresholds — that have made unified registration models hard to build. The ablation isolating normalization, overlap scoring, and the local transformer gives a clean decomposition of where the gains come from.

Real-world applications:

• Mobile robots and drones that move between indoor rooms and outdoor spaces within a single session, where a fixed voxel size or tuned inlier threshold would fail at the transition.
• Autonomous driving and mapping, where LiDAR scans must be aligned across vehicles with different beam counts and against pre-built maps.
• Augmented and mixed reality, where a headset or handheld device registers live depth data against a stored scene model.
• 3D reconstruction and asset pipelines, where objects scanned by different devices must be aligned without per-object tuning.

Industry relevance. Eliminating per-domain hyperparameters matters commercially: a deployed system that changes environments cannot afford an engineer to set the voxel size and inlier threshold for each new site. The roughly O(Nkd) memory scaling also makes a single architecture viable for both small objects and large outdoor scenes.

Future Directions

• Rotation-equivariant backbones. The authors note that the KPConv-based backbone is still rotation-variant and may struggle under extreme rotations; existing rotation-equivariant alternatives were assessed but their memory footprint prevents processing large-scale point clouds in a unified architecture.
• Scaling to very large scenes. RAM may not handle urban-scale UAV captures; tiling scenes into overlapping chunks with streaming is suggested as a practical remedy.
• Robustness of voxel-size selection to outliers. Distant outliers can inflate the initial voxel size and oversmooth geometry; adding a separate outlier-removal step is proposed.
• Broader cross-source evaluation and foundation models. Cross-source registration is discussed only in the supplementary material, and the authors position RAM as a step toward general-purpose point cloud foundation models.

Target Audience

Researchers and graduate students in 3D vision, SLAM, and robotics who work on point cloud registration or on generalizable/foundation models for geometric data; engineers deploying registration in multi-environment systems who need one model rather than per-domain tuning; and practitioners evaluating whether cross-domain training can replace domain adaptation for their sensor mix. Readers wanting only the practical takeaway can focus on the benchmark tables and the normalization ablation.

Authors’ abstract

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.