3DGS_LSR:Large_Scale Relocation for Autonomous Driving Based on 3D Gaussian Splatting

Bo Xu; Haijier Chen; Haitao Lu; Haoze Liu; Shoujian Zhang; Ziao Liu

REVIEW 4 major objections 6 minor 3 cited by

3DGS_LSR:Large_Scale Relocation for Autonomous Driving Based on 3D Gaussian Splatting

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 3DGS-LSR claims that an iterative render-and-refine loop on a 3D Gaussian Splatting map localizes a vehicle from a single monocular image to centimeter-level accuracy in urban driving scenes.

desk verdict The 3DGS-LSR pipeline is a plausible assembly of known components, but the missing train/test split makes the reported centimeter-level accuracy an artifact of memorization rather than a validated relocalization result. read the letter →

arxiv 2507.05661 v1 pith:TAVA3SS6 submitted 2025-07-08 cs.RO cs.CV

Haitao Lu , Haijier Chen , Haoze Liu , Shoujian Zhang , Bo Xu , Ziao Liu This is my paper

classification cs.ROcs.CV

keywords 3DGaussianSplattingvisualrelocalizationautonomousdrivingmonocularcameralocalizationiterativeposerefinementKITTIdatasetfeaturematchingGNSS-free

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes 3DGS-LSR, a relocalization system that uses a prebuilt 3D Gaussian Splatting map of a large outdoor scene to determine a vehicle's pose from a single monocular RGB image, with no GNSS input at run time. The load-bearing claim is that an iterative loop—match the query image to the closest rendered view, solve a camera pose from the matched features, re-render the map from that pose, and repeat until the pose stops changing—converges to centimeter-level positioning accuracy. On KITTI town, boulevard, and highway sequences, the paper reports average positioning errors of 0.026 m, 0.029 m, and 0.081 m, and lower errors than five comparison methods, including a prior 3DGS-based relocalizer. If the claim holds, autonomous vehicles would gain a low-cost, real-time localization channel in the urban environments where GNSS is unreliable.

What carries the argument

The load-bearing component is the 3D Gaussian Splatting map: a set of anisotropic 3D Gaussians with learned color and opacity that can be rasterized into RGB and depth images from any camera pose in milliseconds. Around it runs the refinement loop (Algorithm 3): render at the current pose, extract SuperPoint features and match them to the query with SuperGlue, solve the pose with EPnP plus a least-squares bundle adjustment, and repeat until the pose change falls below 0.01 units. The map does double duty—its rendered RGB views provide the coarse reference gallery, and its rendered depth lets matched pixels be unprojected into 3D—so the same representation carries both the retrieval and the refinement stages.

What would settle it

Rebuild the 3DGS map from the first half of a KITTI sequence and relocalize images from the second half (or from a separate traversal of the same road); if the average error jumps from the reported 2–8 cm to decimeters or meters, the accuracy depends on query frames being part of the map rather than on the render-refine loop.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a 3D Gaussian map can serve not only as a visual reference for retrieval but as an active, renderable pose-refinement device. Starting from a coarse pose obtained by matching the query image to a sparse gallery of rendered views, the method runs a PnP solver on 2D–3D correspondences, then renders a fresh RGB-and-depth pair from the estimated pose and matches again; because each rendered view lands closer to the query, feature matches become more numerous and more confident (the paper reports matches rising from 58 to 383 and average confidence from 0.735 to 0.957 over four iterations), and the pose estimates converge to the reported few-centimeter accuracy. The map itself is built with OmniRe, a panoramic dynamic-scene reconstruction method that fuses LiDAR, camera, IMU, and object masks, which the paper credits for suppressing rendering artifacts from moving vehicles and difficult lighting. On the paper's own numbers, this coarse-to-fine render-refine loop yields average translation errors of 2.9–3.5 cm on KITTI and under 4 cm on Waymo and NuScenes, while requiring only monocular RGB at query time.

Load-bearing premise

The paper does not state whether the query images used in the relocation tests were excluded from the frames used to build the 3DGS map, so the reported centimeter errors could come from the map recognizing its own training views.

Editorial extensions

If this is right

A vehicle equipped with only a standard camera can achieve absolute positioning at the 2–8 cm level in urban driving scenes, once a high-quality 3DGS map of the area has been built.
Relocalization runs in roughly 1–2 seconds per query on an RTX 4090 with about five render-refine iterations, so the pipeline is in reach of near-real-time navigation.
The sparse gallery of rendered reference views suffices even under large viewpoint changes, meaning the map can be stored much more compactly than a dense image database.
The same render-refine pipeline reports average errors under 4 cm on Waymo and NuScenes, indicating the mechanism is not tied to KITTI's specific capture setup.
Feature-match count and confidence rise monotonically across iterations, so the accuracy ceiling is set by rendering fidelity rather than by the quality of the initial retrieval.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If those accuracies survive a true held-out test, 3DGS maps could become the primary absolute-positioning layer in cities, with GNSS and IMU serving only for dead-reckoning between relocalizations.
The render-refine loop is an analysis-by-synthesis strategy; a natural next step is backpropagating through the Gaussian rasterizer to refine the pose directly by image alignment, removing the explicit PnP stage.
The reported accuracy is tied to a map built with LiDAR and object masks, so the centimeter-level claim describes professionally produced maps, not maps reconstructed from a single camera alone.
A decisive extension is season- or time-of-day change: relocalize against a map built months earlier, since the paper's own conclusion notes sensitivity to rendering fidelity under novel viewpoints.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The headline here is that the paper's central claim — centimeter-level pose from a single monocular image via iterative 3DGS rendering — is not supported by the evaluation. The map is built from KITTI camera, LiDAR, IMU, and MASK data from specific sequences, and the query images are selected from those same sequences. No train/test split is mentioned anywhere. So the gallery images and the depth maps used in PnP are generated from geometry that includes the query frames themselves. That makes the reported 0.026/0.027/0.079 m errors look like memorization rather than relocalization.

What is genuinely new here is limited. The pipeline combines OmniRe for mapping, SuperPoint/SuperGlue for matching, EPnP for pose, and the iterative rendering-based refinement that already appeared in iComMa and GSLoc. Applying that to large-scale outdoor driving with a sparse gallery is a legitimate extension, and the paper explains the pieces clearly. Algorithm 3 is concrete, and the timing breakdown (~1–2 seconds per query) is useful. The authors also cite the relevant prior art, including the two iterative-refinement papers, and they acknowledge some limitations in the conclusion.

The soft spots are substantial. The missing train/test split is the load-bearing flaw; without it, the comparison with DSAC++ and 3DGS-Reloc has no evidential force. Table 3 introduces Waymo and NuScenes results with no protocol — no description of map construction or sequence selection. That is unsupported. There is also a minor discrepancy between the abstract (0.029, 0.081) and Table 2 (0.027, 0.079), and no code or data is shared. On the positive side, the engineering is coherent and the writing is honest about limitations, so the problem is not sloppiness in the method itself but in the experimental design.

This paper is for readers interested in seeing a 3DGS relocalization pipeline adapted to outdoor driving, but the numbers should be treated as illustrative at best. I would not cite it in my own work because the results are not established. I would, however, send it to a serious referee: the problem is important, the approach is plausible, and a referee could require a proper hold-out evaluation. As it stands, the paper should be rejected.

Engage with it as a workshop idea, ask for a rigorous train/test split and a transparent Waymo/NuScenes protocol, and don't let the current accuracy claims stand.

Referee Report

4 major / 6 minor

Summary. The paper proposes 3DGS-LSR, a monocular visual relocalization system for autonomous driving. A 3D Gaussian Splatting map is built offline from multi-sensor data (camera, IMU, LiDAR, and object masks), and online localization matches a single RGB query image against rendered reference images using SuperPoint and SuperGlue, estimates an initial pose with PnP, and then iteratively renders new reference views at the estimated pose to refine the result. The authors report centimeter-level positioning accuracy on KITTI (0.026–0.079 m RMSE) and claim superior performance over DSAC++, AS, NG-RANSAC, Regression-only, and 3DGS-Reloc, including additional results on Waymo and NuScenes.

Significance. If the reported accuracy were obtained under a sound evaluation protocol, the system would be a relevant contribution to GNSS-denied urban localization: the use of OmniRe for dynamic outdoor scenes, the sparse gallery rendering, and the rendering-in-the-loop refinement are reasonable design choices. The paper also provides a useful timing breakdown of the localization pipeline. However, the manuscript does not provide code, data, or a clearly specified train/test protocol, and the headline quantitative claims are not supported by the described experiments. The significance can therefore only be assessed after a major re-evaluation.

major comments (4)

[Sections 4.1 and 4.3] The evaluation appears circular. Section 4.1 states that the 3DGS map is built from KITTI camera, IMU, LiDAR, and MASK data from the same sequences, and Section 4.3 selects query images from those same sequences and relocalizes them sequentially, with no train/test split reported. Because the map is optimized against those very frames, and because the depth images used in PnP are rendered from LiDAR points accumulated on the same pass, the initial retrieval, the 2D-3D correspondences, and the iterative refinement all have access to information from the query time. The RMSE values in Table 2 (0.026–0.079 m) are therefore consistent with map memorization and do not establish monocular relocalization accuracy. The authors must either specify and implement a strict temporal or spatial holdout (e.g., map construction from frames strictly preceding the query interval, or from a separate traversal) or remove the quantitative claims.
[Section 4.3 and Table 3] The Waymo and NuScenes results in Table 3 are presented without any experimental protocol. Section 4.1 describes only the KITTI dataset and setup; the manuscript never states how maps are constructed for Waymo and NuScenes, which sequences are used, how query poses and ground truth are obtained, or what split is applied. These columns cannot be verified and should either be fully documented with the same level of detail as the KITTI experiments or removed.
[Abstract, Table 2, and Table 3] The headline numbers are inconsistent. The abstract reports average positioning accuracies of 0.026 m, 0.029 m, and 0.081 m, while Table 2 lists RMSE values of 0.026, 0.027, and 0.079 for the three sequences and mean errors of 0.019, 0.014, and 0.038. Table 3 reports KITTI town-road and boulevard-road accuracies of 2.9 cm and 3.5 cm. The paper should define which error statistic is claimed in the abstract and reconcile the values across Tables 2 and 3.
[Algorithm 3 and Eq. (8)] The iterative refinement, which is the core contribution, is under-specified. The termination condition “If ||P_n − P_{n−1}|| ≤ 0.01” uses a subscript n that is not defined as the loop variable, which is i, so the stopping criterion is ambiguous. In addition, Eq. (8) defines a bundle-adjustment objective but does not specify the projection function Π, the matrices T and K, or the exact pose parameterization, making the refinement step unreproducible.

minor comments (6)

[Conclusion] The text contains typos such as “gasuion splatting” and “rander” in Algorithm 3; these should be corrected.
[References] Reference [37] is the SuperGLUE NLP benchmark paper, not the SuperGlue feature-matching paper by Sarlin et al.; the citation is incorrect and should be replaced with the SuperGlue reference.
[Table 1] The caption “trajectory errors in different scenarios” does not match the table content, which reports iteration statistics; the caption should be changed accordingly.
[Figure 7] Figure 7 is referenced but not described in the text; the axes, the definition of “High-Precision Relocalization Images,” and the threshold used (10 cm and 1 degree) should be stated explicitly.
[Section 4.2] The paper says the reference gallery is rendered sparsely but does not specify the rendering spacing, which is a free parameter that directly affects retrieval difficulty and the reported accuracy.
[Section 4.4] The 1.04–1.84 s total localization time is difficult to reconcile with the claim of “real-time” performance; the paper should state the assumed frame-rate requirement for the target application.

Circularity Check

1 steps flagged · score 7.0 of 10

No train/test split between 3DGS map construction and query images makes the reported centimeter-level localization an artifact of map memorization.

fitted input called prediction [Sec. 4.1; Sec. 4.3; Eq. 5; Algorithm 3]
"we use the camera, IMU inertial navigation device, Velodyne HDL-64E S2 LIDAR sensor data, and the MASK information of the objects to complete the construction of high-precision 3DGS maps of large-scale scenarios, and then we use a single image provided by the camera to complete the relocation ... Then we randomly selected a period of time in each of the three sequences respectively ... and relocated them sequentially as query pictures ..."

The map is explicitly fit to the KITTI camera images and LiDAR depth of the driving sequences (Sec. 4.1; Eq. 5 uses pixel-level L1/SSIM loss against target images and depth loss against LiDAR). Relocalization queries are then taken from the same three sequences (Sec. 4.3), with no train/test split stated. Algorithm 3 renders RGB+depth at the estimated pose from this map and runs PnP between the rendered view and the query; because the query image and its LiDAR depth were already used to optimize the renderer, the 'reference' view at the query pose is a reconstruction of the query itself. The reported 0.026–0.081 m errors therefore measure how well the 3DGS map memorizes the query images, not generalization to unseen viewpoints.

full rationale

The load-bearing empirical claim is the reported centimeter-level accuracy (Abstract, Tables 2–3). The paper constructs the 3DGS map from KITTI camera/IMU/LiDAR/MASK data (Sec. 4.1) and then selects query images from the same three KITTI sequences (Sec. 4.3), without stating any hold-out split. Because the map optimization (Eq. 5) directly minimizes L1/SSIM image differences and LiDAR-depth differences against these very sequences, the rendered RGB/depth images used by Algorithm 3 at the query pose are reconstructions of the query data. The iterative PnP refinement therefore converges to a pose that is consistent with the map's training data, so the reported 0.026 m, 0.029 m, and 0.081 m errors measure self-reconstruction rather than relocalization to unseen views. This is evaluation circularity, not a circular derivation of the algorithm itself; the retrieval->PnP->render loop is a legitimate estimation procedure when the map comes from a separate pass. The self-citations at the end ([45]–[47]) are caveats and are not load-bearing. If the authors provide a proper train/test split or evaluate on map-built-from-prior-pass data, the circularity concern would disappear and the method could be assessed fairly.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method adds no new fitted parameters beyond standard hyperparameters, but it inherits all assumptions of OmniRe map construction and pretrained feature networks. The evaluation assumptions around train/test overlap are not stated.

free parameters (3)

max_i (iteration limit) = 10
Hand-chosen upper limit on refinement iterations in Algorithm 3, affects runtime and convergence.
termination threshold = 0.01 m
Hand-chosen pose change threshold for stopping iteration in Algorithm 3, affects final accuracy.
gallery rendering spacing
The reference image library is rendered sparsely, one image per distance, but the spacing is never specified, which affects retrieval difficulty and accuracy.

assumptions (4)

domain assumption OmniRe produces a geometrically consistent 3DGS map of large-scale outdoor scenes from LiDAR+camera+IMU+mask data.
Section 3.1 relies entirely on OmniRe [28] for map construction; no map quality evaluation is provided.
domain assumption The KITTI ground-truth poses are sufficiently accurate to serve as localization reference.
Section 4.1 uses KITTI ground truth as the positional truth; systematic errors in the dataset reference would bias all reported errors.
domain assumption SuperPoint and SuperGlue generalize from real photographs to 3DGS-rendered images.
Sections 3.2 and 4.2 assume the pre-trained networks produce reliable matches between the query image and rendered reference images, but no fine-tuning or validation on rendered imagery is described.
domain assumption The reference image library is sparse but still contains a reference image close enough in viewpoint for global matching to find a coarse pose.
Section 4.2 states the gallery is rendered sparsely, but no spacing or coverage analysis is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of 3DGS_LSR:Large_Scale Relocation for Autonomous Driving Based on 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/TAVA3SS6

@misc{pith2026250705661,
  author       = {Pith},
  title        = {Pith review of: 3DGS_LSR:Large_Scale Relocation for Autonomous Driving Based on 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TAVA3SS6}},
  note         = {Machine review of arXiv:2507.05661}
}

read the original abstract

In autonomous robotic systems, precise localization is a prerequisite for safe navigation. However, in complex urban environments, GNSS positioning often suffers from signal occlusion and multipath effects, leading to unreliable absolute positioning. Traditional mapping approaches are constrained by storage requirements and computational inefficiency, limiting their applicability to resource-constrained robotic platforms. To address these challenges, we propose 3DGS-LSR: a large-scale relocalization framework leveraging 3D Gaussian Splatting (3DGS), enabling centimeter-level positioning using only a single monocular RGB image on the client side. We combine multi-sensor data to construct high-accuracy 3DGS maps in large outdoor scenes, while the robot-side localization requires just a standard camera input. Using SuperPoint and SuperGlue for feature extraction and matching, our core innovation is an iterative optimization strategy that refines localization results through step-by-step rendering, making it suitable for real-time autonomous navigation. Experimental validation on the KITTI dataset demonstrates our 3DGS-LSR achieves average positioning accuracies of 0.026m, 0.029m, and 0.081m in town roads, boulevard roads, and traffic-dense highways respectively, significantly outperforming other representative methods while requiring only monocular RGB input. This approach provides autonomous robots with reliable localization capabilities even in challenging urban environments where GNSS fails.

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Satellite-Free Training for Drone-View Geo-Localization
cs.CV 2026-04 conditional novelty 7.0 of 10

A satellite-free training framework reconstructs 3D drone scenes via Gaussian splatting, generates geometry-normalized pseudo-orthophotos, and aggregates DINOv3 features with a Fisher vector model trained only on dron...
Satellite-Free Training for Drone-View Geo-Localization
cs.CV 2026-04 conditional novelty 6.0 of 10

A satellite-free pipeline reconstructs multi-view drone scenes with 3D Gaussian splatting, renders PCA-guided pseudo-orthophotos, and learns Fisher-vector retrieval features only from drone data that still match satel...
LSGS-Loc: Towards Robust 3DGS-Based Visual Localization for Large-Scale UAV Scenarios
cs.CV 2026-04 unverdicted novelty 5.0 of 10

LSGS-Loc delivers state-of-the-art accuracy and robustness for 3DGS-based visual localization in large UAV scenes via scale-aware initialization and reliability masking without scene-specific training.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages · cited by 2 Pith papers

[1]

SimultaneousLocalizationandMapping:ASurveyof CurrentTrendsinAutonomousDriving.IEEETransactions onIntelligentVehicles,2,194-220

[1]Bresson,G.,Alsayed,Z.,Yu,L.,&Glaser,S.(2017). SimultaneousLocalizationandMapping:ASurveyof CurrentTrendsinAutonomousDriving.IEEETransactions onIntelligentVehicles,2,194-220. [2]Chalvatzaras,A.,Pratikakis,I.,&Amanatiadis,A.(2023).A SurveyonMap-BasedLocalizationTechniquesfor AutonomousVehicles.IEEETransactionsonIntelligent Vehicles,8,1574-1596. [3]Kuutti...

work page 2017
[79]

3DGS-ReLoc: 3D Gaussian Splatting for Map Representation and Visual ReLocalization

[37]Wang,A.,Pruksachatkun,Y.,Nangia,N.,Singh,A.,Michael, J.,Hill,F.,...&Bowman,S.(2019).Superglue:Astickier benchmarkforgeneral-purposelanguageunderstanding systems.Advancesinneuralinformationprocessing systems,32. [38]Lepetit,V.,Moreno-Noguer,F.,&Fua,P.(2009).EPnP:An accurateO(n)solutiontothePnPproblem.International journalofcomputervision,81,155-166. [3...

work page Pith review arXiv 2019
[6083]

[9]Andrew,Q.(2024).Adaptivebird'seyeviewdescriptionfor long-termmappingandloopclosurein3Dpoint clouds.AppliedandComputationalEngineering

[8]Hu,Y.,Li,S.,Weng,W.,Xu,K.,&Wang,G.(2023).NSAW: AnEfficientandAccurateTransformerforVehicleLiDAR ObjectDetection.IEEETransactionsonInstrumentationand Measurement,72,1-10. [9]Andrew,Q.(2024).Adaptivebird'seyeviewdescriptionfor long-termmappingandloopclosurein3Dpoint clouds.AppliedandComputationalEngineering. [10]Khoo,H.,Ng,Y.,&Tan,C.(2024).OptimizedRecei...

arXiv 2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

SimultaneousLocalizationandMapping:ASurveyof CurrentTrendsinAutonomousDriving.IEEETransactions onIntelligentVehicles,2,194-220

[1]Bresson,G.,Alsayed,Z.,Yu,L.,&Glaser,S.(2017). SimultaneousLocalizationandMapping:ASurveyof CurrentTrendsinAutonomousDriving.IEEETransactions onIntelligentVehicles,2,194-220. [2]Chalvatzaras,A.,Pratikakis,I.,&Amanatiadis,A.(2023).A SurveyonMap-BasedLocalizationTechniquesfor AutonomousVehicles.IEEETransactionsonIntelligent Vehicles,8,1574-1596. [3]Kuutti...

work page 2017

[2] [79]

3DGS-ReLoc: 3D Gaussian Splatting for Map Representation and Visual ReLocalization

[37]Wang,A.,Pruksachatkun,Y.,Nangia,N.,Singh,A.,Michael, J.,Hill,F.,...&Bowman,S.(2019).Superglue:Astickier benchmarkforgeneral-purposelanguageunderstanding systems.Advancesinneuralinformationprocessing systems,32. [38]Lepetit,V.,Moreno-Noguer,F.,&Fua,P.(2009).EPnP:An accurateO(n)solutiontothePnPproblem.International journalofcomputervision,81,155-166. [3...

work page Pith review arXiv 2019

[3] [6083]

[9]Andrew,Q.(2024).Adaptivebird'seyeviewdescriptionfor long-termmappingandloopclosurein3Dpoint clouds.AppliedandComputationalEngineering

[8]Hu,Y.,Li,S.,Weng,W.,Xu,K.,&Wang,G.(2023).NSAW: AnEfficientandAccurateTransformerforVehicleLiDAR ObjectDetection.IEEETransactionsonInstrumentationand Measurement,72,1-10. [9]Andrew,Q.(2024).Adaptivebird'seyeviewdescriptionfor long-termmappingandloopclosurein3Dpoint clouds.AppliedandComputationalEngineering. [10]Khoo,H.,Ng,Y.,&Tan,C.(2024).OptimizedRecei...

arXiv 2023