acceptodds
Withdrawn from ICLR 2027

From Topology to Metric: Geometry-Grounded Refinement for Metric Place Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual Place Recognition (VPR) aims to localize query images by retrieving visually matched geo-tagged database images. Existing works mainly achieve Topological Place Recognition, where the coordinate of the retrieved database image is directly assigned to the query location, overlooking the spatial offset between matched images and only getting the coarse geographical location. To address this issue, we introduce Metric Place Recognition and propose GeoMPR, a geometry-grounded framework that extends topological retrieval toward precise geographic position estimation. GeoMPR first retrieves a top- set of geo-tagged candidates and filters the candidates to form a locally coherent reference set. Given the query and these references, VGGT estimates their relative camera geometry. The Spatial Offset Estimation Module (SOEP) first establishes the transformation mapping between the predicted relative camera geometry and the spatial offset using these reference images, and subsequently leverages this mapping to derive the spatial offset estimation of the query image relative to the anchor, and refines the estimation using residual compensation and global correction. Moreover, we introduce a Geometry-aware Spatial Offset Loss to explicitly supervise the spatial displacement between query and anchor images for accurate geographic position estimation. Extensive experiments on several challenging VPR benchmarks demonstrate that GeoMPR substantially improves place recognition accuracy under strict distance thresholds. Notably, GeoMPR improves the 5-meter recognition accuracy of the SOTA method on Tokyo24/7 by 42.5 pp (36.8% 79.3%), demonstrating the effectiveness of geometry-grounded refinement. Codes will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Topological Place Recognition, Metric Place Recognition, feed-forward neural network

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TVnHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoMPR, a geometry-grounded framework that extends conventional Visual Place Recognition (VPR) from image retrieval to precise geographic localization. It leverages VGGT to estimate relative camera geometry and introduces a Spatial Offset Estimation Module (SOEP) to predict geographic offsets through weighted ridge regression and residual refinement. A geometry-aware loss is also introduced to improve localization accuracy. Experiments on five VPR benchmarks demonstrate substantial improvements over retrieval-based methods, particularly under strict distance thresholds.

Critical strengths
  1. Well-motivated problem: The paper addresses an important limitation of retrieval-based VPR by explicitly estimating spatial offsets rather than directly assigning retrieved image coordinates, enabling more precise geographic localization.

  2. Experimental results: GeoMPR achieves substantial improvements under strict localization thresholds across multiple benchmarks. Its consistent gains with different VPR backbones further demonstrate the effectiveness and generalizability of the proposed framework.

Critical weaknesses
  1. Insufficient Novelty Compared with Existing Coarse-to-Fine Localization Frameworks. The idea of refining an initial retrieval result into a precise location through a second localization stage has been explored in prior works, such as Text2Loc (https://arxiv.org/abs/2311.15977) and MambaPlace (https://arxiv.org/abs/2408.15740). Although these methods focus on different modalities, the paper does not sufficiently discuss their conceptual similarities or clearly distinguish its contributions from existing coarse-to-fine localization frameworks.

  2. Lack of Comparisons with Simple Geometry-Based Baselines. GeoMPR employs VGGT for geometric estimation, followed by weighted ridge regression, residual compensation, and global correction. However, the paper lacks comparisons with simpler alternatives, such as VGGT combined with least squares, standard ridge regression, or appropriate geometric alignment methods (e.g., Procrustes/Umeyama-based alignment). Without these comparisons, it is unclear whether the substantial improvements mainly come from VGGT or the proposed SOEP module.

  3. Limited Methodological Novelty. The proposed framework primarily combines existing techniques, including VGGT, weighted ridge regression, Transformers, and residual refinement. While their integration is reasonable, the paper lacks a sufficiently distinctive algorithmic contribution beyond combining established components.

  4. Missing Ablation Study on the Proposed Geometry-aware Loss. Although Table 5 evaluates the contributions of residual compensation and global correction, the paper does not independently examine the effectiveness of the proposed Geometry-aware Spatial Offset Loss. Comparisons using only Smooth L1 Loss, only Euclidean Distance Loss, and their combination would help establish whether the proposed objective provides meaningful additional benefits.

  5. Lack of Computational Efficiency Analysis. GeoMPR requires an additional VGGT-based geometric inference stage involving the query and multiple reference images, potentially introducing substantial computational overhead. However, the paper does not report inference latency, GPU memory consumption, or computational costs. This makes it difficult to assess the practicality of the proposed framework in large-scale or real-time localization applications.

Questions for authors
  1. Novelty Compared with Existing Coarse-to-Fine Localization Frameworks: How does GeoMPR differ from existing coarse-to-fine localization methods, such as Text2Loc (https://arxiv.org/abs/2311.15977) and MambaPlace (https://arxiv.org/abs/2408.15740)? Could the authors clarify the key conceptual differences and novel contributions of GeoMPR compared with these approaches?

  2. Comparisons with Simple Geometry-Based Baselines: Can the authors provide comparisons with simpler baselines, such as VGGT combined with least squares, standard ridge regression, or appropriate geometric alignment methods? These comparisons would help determine whether the performance improvements primarily originate from VGGT or the proposed SOEP module.

  3. Methodological Novelty: Since GeoMPR mainly integrates existing techniques, including VGGT, weighted ridge regression, Transformers, and residual refinement, could the authors clarify which components represent the main methodological innovations and how they differ from existing approaches?

  4. Ablation Study on Geometry-aware Loss: Can the authors provide an ablation study comparing Smooth L1 Loss alone, Euclidean Distance Loss alone, and their combination? This would help determine whether the proposed Geometry-aware Spatial Offset Loss provides meaningful improvements over standard regression objectives.

  5. Computational Efficiency: Can the authors report the inference latency, GPU memory consumption, and computational costs of GeoMPR compared with conventional VPR methods? Given the additional VGGT-based geometric inference stage, how does the computational overhead affect the practicality of the proposed framework?

Reviewer_VmnwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoMPR refines retrieval-based place recognition into a continuous geographic position estimate. It uses VGGT to recover relative camera geometry, aligns it with reference coordinates through learned weighted regression, and applies residual corrections. Results across five benchmarks show substantial gains at tight localization thresholds, particularly on Tokyo24/7 and SVOX.

Critical strengths

The paper addresses a clear limitation of retrieval-only localization: a visually correct reference can still be several meters away from the query. The improvements at tight distance thresholds are substantial.

The refinement works with several retrieval backbones, suggesting that its benefits extend beyond a single VPR model. The component ablations also support the usefulness of the residual corrections.

Critical weaknesses

The comparisons mainly use retrieval methods that return a database coordinate. They establish the benefit of continuous refinement, but do not adequately establish the value of SOEP. A direct VGGT alignment baseline with uniform weights, together with coordinate interpolation, would help separate the gains from geometry itself and the proposed learned components.

Important inference settings are unspecified, including the retrieval top-k, geographic filtering radius, and handling of too few or poorly distributed references. These directly affect the geometry-to-coordinate mapping, yet their sensitivity is not evaluated. Runtime and memory costs are also unreported despite adding multi-image VGGT inference.

The novelty of Metric Place Recognition is a bit overstated without sufficient discussion of some prior retrieval-based localization and pose refinement methods [1][2][3][4]. Although their assumptions differ, these are relevant to GeoMPR’s reference selection and geometric estimation, and their connections and differences should be discussed.

[1] Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features. CVPR 2025

[2] Doppelgangers: Learning to Disambiguate Images of Similar Structures. ICCV 2023

[3] AIR-HLoc: Adaptive Retrieved Images Selection for Efficient Visual Localisation. ICRA 2025

[4] GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting. ICLR 2025

Questions for authors

How does GeoMPR compare with uniform-weight ridge alignment using the same VGGT outputs and references, without learned corrections?

What top-k and filtering radius are used at inference, and how does performance change when the retained references are sparse, nearly collinear, or share the same coordinates?

Reviewer_rEFaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  1. Method. The paper proposes GeoMPR, a post-retrieval refinement framework for estimating a query image's 2D geographic coordinate rather than assigning it the coordinate of the top-retrieved database image.

  2. Pipeline. A conventional VPR model retrieves top-kk geo-tagged candidates. Candidates near the top-1 anchor are retained, and VGGT jointly predicts local camera centers for the query and references. The Spatial Offset Estimation Module (SOEP) fits a learned-weight differentiable ridge-regression map from the reference camera centers to their known anchor-relative UTM offsets, applies it to the query center, and adds learned residual-interpolation and global-correction terms. Training uses a Smooth-L1 offset loss plus a Euclidean position-error term.

  3. Validation. GeoMPR is trained on GSV-Cities and evaluated on Pitts30k, MSLS-val, Tokyo24/7, SVOX, and Baidu Mall. Relative to assigning the top-1 retrieval coordinate, it reports substantial improvements at strict distance thresholds, especially on Tokyo24/7 and SVOX. Experiments also vary the retrieval backbone, VGGT fine-tuning, and the two correction terms.

Critical strengths
  1. Useful correction of a genuine retrieval limitation. Returning a database coordinate imposes an error floor determined by sampling density. GeoMPR directly estimates the missing spatial offset and produces particularly large improvements under strict 3 m and 5 m criteria. Improvements across multiple retrieval backbones suggest that the refinement is not tied only to ImAge.

  2. A flexible weak-map setting. Unlike many classical visual-localization systems, GeoMPR requires only reference images and 2D geo-tags rather than an offline SfM point cloud, calibrated reference intrinsics, or full 6-DoF database poses. This makes the approach potentially useful for existing geo-tagged VPR databases.

Critical weaknesses
  1. The novelty claim overlooks near-direct predecessors. The paper presents “Metric Place Recognition” as a new formulation, but retrieval followed by continuous metric position estimation is established, including almost the same abstract pipeline. Sattler et al., Are Large-Scale 3D Models Really Necessary for Accurate Visual Localization? (CVPR 2017), retrieve candidates, retain images within 25 m of the nearest-neighbor or re-ranked anchor, jointly reconstruct those references and the query with local SfM, register the reconstructed camera centers to UTM using reference geo-tags, and output the query's global metric position. Salarian et al., Improved Image-Based Localization Using SFM and Modified Coordinate System Transfer (IEEE TMM 2018), select multiple retrieved GPS-tagged references, jointly reconstruct their and the query's camera centers, fit a similarity transformation from reconstruction coordinates to geographic coordinates, and apply it to estimate the query geo-tag. Salarian et al. can additionally use a coarse noisy query-GPS prior, so its operating setup is not identical. Nevertheless, both works implement the claimed topology-to-metric transition using a geo-tagged reference database. Replacing SfM and similarity registration with VGGT, weighted ridge regression, and learned corrections is a potentially useful modernization, but it is a much narrower contribution than claimed. Modern systems such as DUSt3R localization, Reloc3r, Reloc-VGGT, and the concurrent RIC-Loc preprint provide additional context, although they require reference 3D information, calibration, or full poses and therefore are not equal-input baselines. GeoMPR discusses DUSt3R as a geometry foundation model but not its localization protocol; the two closer classical geo-tag-based predecessors are not discussed or compared. Finally, the “Geometry-aware Spatial Offset Loss” is a standard combination of Smooth-L1 coordinate regression and Euclidean coordinate error; it is metric-distance-aware but threshold-agnostic and contains no separate camera-geometric constraint.

  2. The evaluation compares mainly against methods that do not solve the proposed task and lacks the decisive controls. All headline baselines are retrieval-only systems that are forced to output a database coordinate, so gains from any continuous interpolation or geometric localization method are expected. The most directly matched missing baselines are Sattler-style local SfM plus UTM registration, Salarian-style SfM plus coordinate transfer, simple top-kk coordinate interpolation, and VGGT camera centers with standard geometric alignment. Local matching/PnP and modern DUSt3R/MASt3R/VGGT localization would be useful contextual comparisons, but their additional map, pose, or calibration assumptions must be stated explicitly. Even within GeoMPR, the paper omits the key ablations: VGGT centers plus ordinary least squares; similarity/Umeyama alignment; robust regression/RANSAC; weighted ridge without the Transformer; ridge-only and global-correction-only models; loss and scale-normalization ablations; and direct learned offset regression without VGGT. Table 5 removes only the two final corrections. Table 4 is intended to isolate VGGT adaptation, but its “w/o training” label does not state whether SOEP remains trained and all other components are held identical. Consequently, the experiments establish that the full pipeline improves a deliberately coarse top-1 baseline, but not that the proposed SOEP, loss, or claimed formulation is the reason, nor that it advances prior geo-tag-only metric localization.

  3. Practical evaluation and data controls are incomplete. The paper provides no runtime/memory reporting for a VGGT pass over each query and reference set, error CDF or mean/median localization error, uncertainty over the small Tokyo24/7 and MSLS-val query sets, or analysis by reference count, spatial spread, retrieval failure, or VGGT conditioning. It also does not audit geographic or image overlap between GSV-Cities training data and evaluation sets that use Google Street View imagery. These omissions matter because the method's cost, generalization, and behavior under sparse, collinear, or incorrect retrieval neighborhoods determine its practical value.

Questions for authors
  1. How does GeoMPR compare, under matched inputs and reference sets, with (i) coordinate/pose interpolation, (ii) unweighted ridge regression, (iii) similarity/Umeyama alignment, and (iv) robust alignment such as RANSAC? How much do the Transformer, learned weights, residual interpolation, global correction, and the second loss term each contribute beyond these controls?

  2. How is GeoMPR technically and empirically distinguished from Sattler et al.'s retrieval + proximity-filtered local SfM + UTM registration (CVPR 2017) and Salarian et al.'s retrieval + joint SfM + geographic coordinate transfer (TMM 2018)? These use geo-tagged reference images and nearly the same high-level pipeline. A comparison with their classical reconstruction/alignment approach would materially affect my novelty assessment; comparisons with modern full-pose or map-based localization should separately identify their stronger input assumptions.

  3. Since p0=0p_0=0, how is cos⁡(pq,p0)\cos(p_q,p_0) in Eq. (3) computed without division by zero? What happens when geographic filtering retains only the anchor, too few informative references, or a nearly collinear set? Please report the minimum NN, fallback behavior, conditioning statistics, and accuracy as a function of NN and reference geometry.

  4. What are the exact inference values of top-kk and τ\tau, all SOEP dimensions and constants (λfloor,sres,sc,β,λmetric\lambda_{\mathrm{floor}}, s_{\mathrm{res}}, s_c,\beta,\lambda_{\mathrm{metric}}), image resolution, and per-query runtime/memory? In Table 4, is “w/o training” a frozen pretrained VGGT with a trained SOEP, or is no component trained?

  5. Why are training references selected using distance to the ground-truth query location while inference references are selected around the retrieved anchor? What happens when training instead uses the inference-time construction? Please also document checks excluding geographic, panorama-level, or image overlap between GSV-Cities training data and all evaluation sets.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.