PaperlayerView exact revision on arXiv

Source-linked AI summary

ReLoc: Rethinking Scene Coordinate Regression Architecture for Robust Outdoor LiDAR-based Localization

Heejoon Moon, Yurim Cho, Je Hyeong Hong

arXiv:2609.33344v1cs.CV

TL;DR

Existing SCR methods have limited fine-grained global discrimination and include noisy dynamic or unstable local features. ReLoc combines context-aware global embeddings with attention-based local enhancement; it reports state-of-the-art accuracy on two outdoor datasets while maintaining real-time inference.

  • Problem

    Existing SCR methods struggle to distinguish nearby locations with classification-based global embeddings and uniformly sampled local features can include noisy dynamic or unstable regions.

    Source: §I · p. 1

  • Method

    ReLoc combines learnable context tokens and a feature aggregator for global context with attention-based local feature enhancement.

    Source: §I · pp. 1–2

  • Results

    ReLoc reduces mean positional and rotational errors by 15% and 13% on average compared to GTRLoc, while maintaining real-time performance.

    Source: §I · p. 2, B. Localization results · p. 6

  • Takeaways & Limitations

    On two large-scale outdoor datasets, the two modules work synergistically to improve LiDAR localization accuracy beyond prior SCR methods while maintaining real-time performance.

    Source: §I · p. 2

Abstract

from arXiv · show

Scene Coordinate Regression (SCR) has recently emerged as a promising approach for LiDAR-based localization, achieving accurate localization without requiring an explicit 3D map. Despite their effectiveness, existing SCR methods rely on scene classification-based global embedding that struggles to provide fine-grained discrimination among nearby locations. Moreover, their reliance on uniform sampling of local features during training assigns equal importance to all points, thereby inadvertently propagating features from dynamic objects or unstable regions and potentially degrading training stability. In this paper, we present ReLoc, a revamped SCR architecture that can effectively address these limitations. First, we redesign the global embedding module by combining learnable context tokens with a feature aggregator to capture richer and more discriminative scene context. Second, we introduce an attention-based local feature enhancement module to mitigate the impact of noisy local features while encouraging context-consistent structures, yielding more robust local feature representations. Experimental results on two large-scale outdoor datasets demonstrate that our approach achieves state-of-the-art accuracy over previous SCR-based methods while maintaining real-time inference performance.

I. INTRODUCTION

Map-based localization can be accurate but brings storage, communication, and privacy costs. ReLoc targets weaknesses in SCR global and local representations with context-aware global embeddings and enhanced local features.

I. INTRODUCTION

  • Explicit-map localization stores 3D points and descriptors, creating substantial storage demands, communication overhead, and privacy concerns during transmission.SCR instead predicts per-point correspondences to world coordinates and uses robust solvers such as RANSAC to estimate pose.Source: §I · p. 1
  • Recent SCR methods, use classification-based global embeddings that lack fine-grained discrimination among locations within the same scene cluster.Uniform local-feature sampling also includes noisy features from dynamic objects or unstable regions, which can adversely affect training stability and localization.Source: §I · p. 1
  • ReLoc combines learnable context tokens with an MLP-Mixer-based aggregator to form a compact, discriminative global scene embedding.The tokens represent richer scene context, addressing the limited within-cluster discrimination of classification-based embeddings.Source: §I · pp. 1–2
  • ReLoc’s attention-based local enhancement refines features through residual connections to mitigate noise and promote context-consistent interactions.Experiments on two large-scale outdoor datasets show improved accuracy over prior SCR methods while maintaining real-time performance.Source: §I · p. 2

II. RELATED WORK

Prior LiDAR localization work includes explicit-map retrieval and registration, as well as direct pose regression. The related-work discussion contrasts their limitations and positions ReLoc’s context tokens against farthest-point sampling.

A. Structure-based LiDAR localization

  • Structure-based localization uses explicit maps and descriptor matching; retrieval searches for similar scans, while registration estimates poses through matching and optimization.These approaches incur memory overhead from explicit map storage.Source: A. Structure-based LiDAR localization · p. 2

B. Absolute pose regression

  • Absolute pose regression estimates a 6-DoF pose directly from a query point cloud, but APR methods still struggle to achieve meter-level accuracy in complex outdoor environments.Later approaches include spatio-temporal aggregation, bio-inspired memory, Euclidean-hyperbolic feature fusion, and diffusion-based refinement.Source: B. Absolute pose regression · p. 2
  • LightLoc’s dynamic-category features have higher regression errors and lower inlier ratios than features from static regions.The analysis uses a pretrained SphereFormer to categorize query points in a QEOxford sequence.Source: Figure 3 · p. 2
  • Unlike FlashMix, which uses farthest-point sampling that includes unstable dynamic elements, ReLoc uses learnable context tokens to capture scene context.The paper notes that such tokens have been shown to focus on geometrically consistent structures.Source: B. Absolute pose regression · p. 2

C. Scene coordinate regression

SCR predicts world coordinates for LiDAR points and estimates pose from the resulting correspondences. The review identifies limited global discrimination and noisy local features as representational bottlenecks.

C. Scene coordinate regression

  • SCR methods regress world coordinates for query points to form 3D-3D correspondences, then estimate pose by solving Procrustes alignment with RANSAC.Prior methods include SGLoc, LightLoc, and GTRLoc, with different feature extractors and regression designs.Source: C. Scene coordinate regression · p. 2
  • LiSA partly addresses noisy local features through semantic awareness, but its performance relies on prior segmentation quality.Other SCR methods sample local features uniformly, including noisy features from dynamic objects or unstable regions.Source: C. Scene coordinate regression · p. 3

III. REVIEW OF SCENE COORDINATE REGRESSION

  • Given predicted correspondences between scan points and world points, a robust LiDAR-to-world transformation is estimated using a registration solver with RANSAC.The correspondence count is denoted by M.Source: §III · p. 3
  • Classification-based global embeddings can assign high similarity to spatially distant samples, creating ambiguity that hinders accurate coordinate learning.Dynamic objects tend to yield higher regression errors and lower inlier ratios than static categories, motivating more discriminative global context and refined local representations.Source: §III · p. 3

IV. PROPOSED METHOD

ReLoc trains its global embedding module first, then freezes it while training local-feature enhancement and a scene-specific regressor. The resulting global and refined local features are combined to predict world coordinates.

IV. PROPOSED METHOD

  • In Stage 1, the global embedding module learns context-aware representations from learnable context tokens and local features.A pose regressor predicts 6-DoF pose, while a projection head encourages similar embeddings for positive scans captured at the same location in different sequences.Source: §IV · p. 3
  • In Stage 2, the trained global embedding module is fixed while the local enhancement module and scene-specific MLP regressor are trained.The enhancement module refines local features from the scene-agnostic backbone.Source: §IV · p. 3
  • The global representation and refined local features are concatenated as hierarchical embeddings and fed to the MLP regressor to predict world coordinates..Source: §IV · p. 3

A. Scene context-aware global embedding module

The global embedding module extracts scan-specific context with learnable tokens and aggregates it into a compact descriptor. That descriptor supports pose regression and contrastive regularization.

A. Scene context-aware global embedding module

  • Learnable queries attend to voxel features to extract context tokens that preserve sub-region geometric structures.The resulting tokens capture scan-specific context through cross-attention.Source: A. Scene context-aware global… · p. 4
  • An MLP-Mixer alternates token-wise and channel-wise mixing to model relationships across scene regions and refine token features.A linear projection and global average pooling produce one global descriptor per scan.Source: A. Scene context-aware global… · p. 4, Figure 5 · p. 4
  • The global descriptor feeds an MLP pose head that predicts 3D position and orientation in logarithmic quaternion form.A lightweight projection layer also applies contrastive regularization across positive scans.Source: A. Scene context-aware global… · p. 4
  • Global-module training combines pose loss with a weighted Barlow Twins contrastive regularization term.The pose loss sums L1 losses for predicted translation and orientation.Source: A. Scene context-aware global… · p. 4

B. Local feature enhancement module

The local feature enhancement module uses self-attention and residual connections to refine voxel features while retaining backbone geometry. Its regression training uses an L1 loss over predicted points.

B. Local feature enhancement module

  • Self-attention models relationships across voxel features, and a residual connection adds the attended refinement to the original features.The residual formulation preserves backbone geometric cues while attenuating responses from dynamic objects or unstable regions using reliable static context.Source: B. Local feature enhancement module · pp. 4–5
  • The feature-enhancement module and MLP regressor are trained with an L1 loss between predicted and ground-truth point coordinates.The loss sums over M predicted points.Source: B. Local feature enhancement module · p. 5

C. LiDAR localization procedure

The localization procedure processes a voxelized LiDAR scan through a scene-agnostic backbone and embedding modules, then regresses world coordinates. A registration solver with RANSAC estimates the 6-DoF pose.

C. LiDAR localization procedure

  • A voxelized LiDAR scan passes through a scene-agnostic backbone, global embedding module, and local feature enhancement module to form hierarchical features.The representations are concatenated before coordinate regression.Source: C. LiDAR localization procedure · p. 5
  • An MLP regressor maps the hierarchical features to predicted world coordinates for scan points.The predicted coordinates are used to establish point correspondences.Source: C. LiDAR localization procedure · p. 5
  • A point-cloud registration solver with RANSAC estimates the 6-DoF pose using a 2 m inlier threshold.The threshold follows LightLoc.Source: C. LiDAR localization procedure · p. 5

A. Experimental details

ReLoc is evaluated on QEOxford and NCLT using prior-study train/test sequences and mean position and orientation errors. It outperforms GTRLoc on both datasets and shows reduced trajectory position errors qualitatively.

A. Experimental details

  • Evaluation covers QEOxford and NCLT, using the same train/test sequences as prior studies.NCLT comprises 27 campus mapping sessions under varying seasonal and illumination conditions.Source: A. Experimental details · p. 5
  • Comparisons include LiDAR-based SCR methods such as GTRLoc, using mean position and orientation errors between predicted and ground-truth poses.Source: A. Experimental details · p. 5
  • Training uses AdamW with a one-cycle learning-rate schedule, and voxel sizes of 0.25 m for QEOxford and 0.3 m for NCLT.The learning rate varies from 5×10−4 to 5×10−3.Source: A. Experimental details · p. 5

B. Localization results

C. Inference time

ReLoc runs faster than the input frame rates on both evaluated datasets. Ablations report gains from the global embedding and local feature enhancement modules, while hyperparameter tests indicate a balance between context injection and preserving local geometry.

C. Inference time

  • 97Hz on QEOxford and 93Hz on NCLT are ReLoc’s average inference speeds, exceeding the datasets’ 20Hz and 10Hz input frame rates.Source: C. Inference time · p. 6

D. Ablation studies

  • The global embedding improves localization over the baseline on both datasets and provides finer discrimination than classification-based embeddings, including higher accuracy than LightLoc when only the global embedding is replaced.The embeddings yield high similarity for nearby scans and lower similarity for distant scans.Source: D. Ablation studies · p. 6
  • Local feature enhancement improves localization over naive scene-agnostic features and improves median regression error and inlier ratio for static and dynamic-region features.The authors attribute this to reinforcing static features and regularizing noisy dynamic features using surrounding context-consistent geometry.Source: D. Ablation studies · p. 6
  • K = 128 achieves the best observed localization performance without significant gains from increasing token count, while increasing β tends to degrade translation accuracy.Excessive feature enhancement can disrupt the original local geometry, weakening the balance between context injection and feature preservation.Source: D. Ablation studies · pp. 6–7

VI. CONCLUSION

ReLoc addresses limitations in SCR global scene discrimination and sensitivity to dynamic or unstable regions with two complementary modules. The paper reports improved localization on two outdoor datasets without compromising real-time inference.

VI. CONCLUSION

  • Current SCR architectures use scene-classification logits that lack fine-grained scene-position discrimination and are vulnerable to noise from dynamic or unstable objects.Source: §VI · p. 7
  • ReLoc combines learnable context tokens and an MLP-Mixer-based aggregator for global embeddings with self-attention-based local feature enhancement.The local module emphasizes stable static structures while mitigating noisy dynamic features.Source: §VI · p. 7
  • Across two outdoor datasets, the modules synergistically improve LiDAR-based SCR localization without compromising real-time inference.Source: §VI · p. 7
Loading 2609.33344v1…