acceptodds
Under review as a conference paper at ICLR 2027

GeLoc3r: Enhancing Relative Camera Pose Regression with Geometric Consistency Regularization

Abstract

ReLoc3R achieves competitive performance with fast 33ms inference and state-of-the-art regression accuracy, yet our analysis reveals subtle geometric inconsistencies in its internal representations that prevent reaching the precision ceiling of correspondence-based methods like MASt3R (which requires 294ms per pair). In this work, we present GeLoc3r, a novel approach to relative camera pose estimation that enhances pose regression methods through Geometric Consistency Regularization (GCR). GeLoc3r overcomes the speed-accuracy dilemma by training regression networks to produce geometrically consistent poses without inference-time geometric computation. During training, GeLoc3r leverages ground-truth depth to generate dense 3D-2D correspondences, weights them using a FusionTransformer that learns correspondence importance, and computes geometrically-consistent poses via weighted RANSAC. This creates a consistency loss that transfers geometric knowledge into the regression network. Unlike the FAR model, which requires both regression and geometric solving at inference, GeLoc3r only uses the enhanced regression head at test time, maintaining ReLoc3R's fast speed and approaching MASt3R's high accuracy. Against a vanilla ReLoc3R trained on the same data, GeLoc3r improves AUC@ from 34.20% to 41.42% on CO3Dv2 (21% relative), from 63.54% to 68.66% on RealEstate10K, and from 39.86% to 50.45% on MegaDepth1500. By teaching geometric consistency during training rather than enforcing it at inference, GeLoc3r offers a new perspective on how neural networks learn camera geometry, achieving both the speed of regression and the geometric understanding of correspondence methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Relative Pose Estimation, Camera Pose Regression, Visual Localization, 3D Computer Vision

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_bEQROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Main contribution: GeLoc3r adds Geometric Consistency Regularization (GCR) to ReLoc3r during training, while preserving direct relative pose regression at inference.

Method:

  1. Reconstruct 3D points from ground-truth depth and camera intrinsics, then project them into the other image using the predicted pose.
  2. Encourage agreement between MASt3R descriptors at the corresponding locations.
  3. Use FusionTransformer weights for weighted RANSAC sampling and enforce consistency between the regressed and solver-estimated poses.

Reported evidence:

  • Training uses ten datasets; evaluation covers five relative pose benchmarks and two localization benchmarks unseen during training.
  • Against ReLoc3r retrained on the same data, AUC@5° improves from 39.86 to 50.45 on MegaDepth1500 and from 34.20 to 41.42 on CO3Dv2.
  • Both models are reported to run at 33 ms per image pair. The added geometric modules are used only during training.
Critical strengths

1. Improved accuracy without additional geometric computation at inference.

  • Design: Depth, descriptor reprojection, and the geometric solver provide supervision only during training.
  • Evidence: Figures 2–3 preserve the direct pose regression path. Table 1 reports the same runtime of 33 ms per image pair for both models.
  • Significance: This preserves the practical benefit of fast pose regression while improving accuracy.

2. Consistent gains over the baseline retrained on the same data.

  • Relative pose: Table 1 shows improvements at all three AUC thresholds across five benchmarks.
  • Localization: On unseen 7-Scenes, the reported average rotation error improves from 1.12° to 1.03°, while position error remains 0.04 m at the reported precision. On unseen Cambridge, average position and rotation errors improve from 0.60 m / 0.61° to 0.51 m / 0.52° (Tables 2–3).
  • Ablation: Table 4 shows gains when adding the descriptor loss and then the full configuration.
  • Significance: The evidence supports the usefulness of the overall training configuration across multiple datasets and metrics.
Critical weaknesses

1. The main baseline comparison does not sufficiently establish the source and extent of the gains.

What is established:

  • Table 1 compares GeLoc3r with a ReLoc3r baseline retrained on the same data. Some improvements over this baseline are substantial.
  • Table 4 provides supporting evidence under a controlled, smaller setting: all variants use MASt3R initialization and 10,000 ScanNet training pairs, with AUC@5° increasing from 19.4 to 20.3 and then 22.4.

What remains unclear:

  • For Table 1, the paper does not sufficiently specify whether the baseline and GeLoc3r share initialization, training budget, corresponding trainable modules, optimization settings, and model-selection criteria beyond the proposed GCR additions.
  • The retrained baseline is substantially below the published ReLoc3r result on MegaDepth1500: 39.86 versus 49.6 AUC@5°, while GeLoc3r reaches 50.45. The cause of this gap is not explained.
  • The original ReLoc3r uses DUSt3R initialization, whereas GeLoc3r uses MASt3R. The initialization of the main Table 1 baseline needs to be specified separately. ReLoc3r's own Table 8 demonstrates sensitivity to initialization on ScanNet1500, making this a concrete comparison concern rather than evidence of an implementation error.
  • Table 4 supports the benefit of the overall additions, but does not resolve the main baseline's performance gap or isolate the effects of learned weighting and the prior.

Effect on my assessment: The published numbers come from different training conditions and do not invalidate the internal comparison. However, without a sufficiently specified and validated baseline, I cannot confidently determine how much of the main improvement is attributable to GCR or how substantial the advance is over an appropriately trained ReLoc3r model. This uncertainty is the primary reason for my weak rejection; I am not claiming that the baseline was trained incorrectly. Validation of the baseline and convincing gains under controlled conditions could change my recommendation. If the authors validate that the baseline is appropriately trained and evaluated, and demonstrate convincing improvements attributable to GCR under controlled conditions, I would be willing to revise my recommendation toward acceptance.

References: GeLoc3r Section 4.1 and Tables 1 and 4; ReLoc3r, arXiv:2412.08376v2, implementation details and Tables 8 and 11.

2. The construction and corrective effect of the solver supervision remain unclear.

What the paper describes:

  • Section 3.5 defines descriptor pairs at the same subsampled grid index. How these pairs and their weights map to the PnP correspondences is not clearly explained.
  • Figure 3's caption describes constructing 3D–2D correspondences by transforming and projecting depth-derived points using the current regressed pose.
  • It does not specify a step that corrects those projected locations through image-based descriptor matching.
  • Figure 3 labels the 3D points in the cam2 frame, leaving the coordinate convention for the PnP inputs unclear.

Why this matters:

  • If source-frame points XiX_i are paired with target coordinates vi=π(K2PregXi)v_i=\pi(K_2P_{\mathrm{reg}}X_i) and passed directly to PnP, then PregP_{\mathrm{reg}} itself has zero reprojection error.
  • Under this construction, changing sampling weights alone does not introduce an independent observation that corrects the pose.
  • Table 4 shows improved final-model performance, but does not directly show that the solver output is closer to ground truth than its input.

Effect on my assessment: I cannot adequately assess the central supervision mechanism from the current description and evidence. This is a conditional concern about the documented construction; it does not establish that the implemented solver is ineffective.

References: Sections 3.5–3.6, Figure 3, Algorithm 1, Table 4.

Questions for authors

1. Validate the Table 1 baseline and clarify the attribution of the gains.

  • Does the baseline use the original DUSt3R initialization or the MASt3R initialization used by GeLoc3r? Please specify all differences between the two models beyond GCR, including training budget, corresponding trainable modules, optimization settings, and model-selection criteria.
  • What explains the baseline's lower MegaDepth1500 result relative to published ReLoc3r? If the original checkpoint has been evaluated with your pipeline, please report that result to help distinguish evaluation differences from retraining differences.
  • What evidence establishes that the baseline is appropriately trained and that the gains persist when these conditions are controlled? This is central to whether I would reconsider my recommendation.

2. Define the correspondences used by PnP.

  • Specify the coordinate frame of the 3D points and how target 2D coordinates are obtained.
  • Section 3.5 defines descriptor pairs at the same subsampled grid index, whereas Figure 3 describes pose-projected coordinates. How do these descriptor pairs and their weights map to the 3D–2D correspondences passed to PnP?
  • Are the target 2D coordinates those grid locations, projections under the predicted pose, or independently estimated image correspondences? If projections alone are used, what supplies a corrective signal beyond recovering the pose that generated them?

3. What evidence shows that the solver improves its input pose?

  • Can you provide ground-truth rotation and translation-direction errors for PregP_{\mathrm{reg}} and PsolverP_{\mathrm{solver}} on the same training pairs, including the fraction improved and the magnitudes of improvement and degradation?
  • This evidence would directly test the current claim that the solver provides more accurate supervision.

4. Clarify and isolate the changes in the ablation.

  • Does the “+ Descriptor loss” variant in Table 4 also learn FusionTransformer weights? What exactly changes in the final row?
  • If existing comparisons with uniform weighting or without the prior are available, please share them to clarify which components account for the gain.
  • Appendix A.1 states that β\beta and τ\tau were selected through ablation studies. What did those studies show?
5 comments

We thank the reviewer for the careful reading and the constructive questions. We answer each question in turn.

Q1. Validity of the Table 1 baseline and attribution of the gains

Q1a. Does the baseline use the original DUSt3R initialization or the MASt3R initialization used by GeLoc3r? Please specify all differences between the two models beyond GCR, including training budget, corresponding trainable modules, optimization settings, and model-selection criteria.

The Table 1 baseline uses the same MASt3R 512-DPT initialization as GeLoc3r, not the original DUSt3R initialization. We chose MASt3R because GCR relies on MASt3R's descriptor heads, which must be compatible with the backbone features. To keep the comparison controlled, the baseline uses the same initialization.

Apart from the GCR components, the two models use the same training configuration:

SettingReLoc3R-512 (retrained baseline)GeLoc3r-512
InitializationMASt3R 512-DPTMASt3R 512-DPT
Frozen modulesViT encoderViT encoder, MASt3R descriptor heads
Trainable modulesdecoder, pose headdecoder, pose head, FusionTransformer (training only)
Training data10 datasets, 50K pairs per dataset per epochsame
Training budget100 epochssame
Batch size / hardware10 pairs per GPU on 8× L40Ssame
OptimizerAdamW, LR 10−5→10−710^{-5} \rightarrow 10^{-7} cosine, gradient clipping 1.0same
LossLpose\mathcal{L}_{pose}λposeLpose+λconsistencyLconsistency+λdescLdescriptor\lambda_{pose}\mathcal{L}_{pose} + \lambda_{consistency}\mathcal{L}_{consistency} + \lambda_{desc}\mathcal{L}_{descriptor} (0.8 / 0.1 / 0.1)
Model selectionLowest median evaluation losssame
Inference architecture and runtimeencoder–decoder–pose head, 33 msidentical, 33 ms

Using three evaluation datasets same as the official reloc3r evaluation protocol, we evaluated both models after each epoch on 1,000 pairs each from ScanNet1500, the ARKitScenes test split, and the MegaDepth test split. For each model, we selected the checkpoint that achieved the lowest median evaluation loss on any of the three datasets. ScanNet1500 was used for both checkpoint selection and final benchmark evaluation.

The FusionTransformer, the descriptor heads and weighted RANSAC are used only during training, so the two models are architecturally identical at test time.

Q1b. What explains the baseline's lower MegaDepth1500 result relative to published ReLoc3R? If the original checkpoint has been evaluated with your pipeline, please report that result.

We evaluated the official ReLoc3R-512 checkpoint with our evaluation pipeline and reproduce the published results (MegaDepth1500: 49.6/67.9/81.2 AUC@5°/10°/20°; ScanNet1500: 34.79/58.37/75.56; RealEstate10K: 66.70/80.20/88.39). The gap therefore comes from our retraining, not from evaluation differences.

Our retraining differs from the official recipe in three respects:

  1. Initialization. We use MASt3R instead of DUSt3R because GCR requires compatible MASt3R descriptor heads. ReLoc3R's own Table 8 shows that this change alone lowers ScanNet1500 AUC@5° from 34.79 to 32.62. Our baseline (33.57) is consistent with that figure.
  2. Training data composition and scale. Official ReLoc3R is trained on about 8M pairs from seven datasets: CO3Dv2, ScanNet++, ARKitScenes, BlendedMVS, MegaDepth (about 1.8M pairs), DL3DV (about 1.1M pairs) and RealEstate10K. Our 10-dataset mixture shares six of them and, in place of DL3DV, adds StaticThings3D, Waymo Open, WildRGB-D and Habitat. We do not use DL3DV, and the official repository notes that its released training setup is “not strictly equivalent to what was used to train Reloc3r,” so the official mixture cannot be reproduced exactly. Both recipes sample 50K pairs from each dataset per epoch, so MegaDepth makes up one tenth of each of our epochs, against one seventh in the official mixture, which plausibly matters most for MegaDepth1500's large viewpoint and intrinsics variations.
  3. Training budget. The official ReLoc3R script trains for 100 epochs with 5 warm-up epochs, a learning rate of 10−510^{-5} decaying to 10−710^{-7}, a frozen encoder, and 50K pairs per dataset per epoch. GeLoc3r and our retrained baseline each train for 100 epochs with 10 pairs per GPU on eight L40S GPUs (80 pairs per batch across GPUs). The released script uses 32 pairs per GPU on eight H800 GPUs (256 pairs per batch across GPUs). The official paper reports a batch size of 8 on eight AMD MI250x-40G GPUs, and the repository states that the released script is “not strictly equivalent to what was used to train Reloc3r”. We therefore cannot match the official budget exactly.

The gap to the released model therefore reflects the differences in training recipe listed above, not the evaluation. Our retrained baseline is trained under the settings listed in Q1a and serves as the reference for Table 1. The evidence that the gains persist under controlled conditions is given in Q1c.

Q1c. What evidence establishes that the baseline is appropriately trained and that the gains persist when these conditions are controlled? This is central to whether I would reconsider my recommendation.

We address the two parts in turn.

(a) The baseline is appropriately trained.

  • Evaluation is validated. Our pipeline reproduces the published results of the official ReLoc3R-512 checkpoint, so any baseline gap reflects training, not evaluation.
  • Matched training budget. Both models train for 100 epochs with the same data sampling and batch size, as specified in Q1a and Appendix A.2.
  • Low run-to-run variance. All results are averaged over three runs. The standard deviation of AUC@5° is ±0.1 for the baseline and ±0.12 for GeLoc3r, well below the reported gains (e.g., +10.6 on MegaDepth1500).

(b) The gains persist under controlled conditions.

  • Controlled comparison at full scale (Table 1). Table 1 compares GeLoc3r with the retrained baseline under the settings listed in Q1a. GeLoc3r improves on all five relative-pose benchmarks and both localization benchmarks.
  • Small-scale ablation (Table 4). On ScanNet1500, the pose-only, descriptor-loss, and full variants achieved 19.4, 20.3, and 22.4 AUC@5°, respectively.

Q2. Correspondences used by PnP

We thank the reviewer for pointing out this ambiguity. Figure 3 and Section 3.5 of the revised paper explain how the correspondences passed to PnP are constructed.

(1) Coordinate frame and 2D targets

  • 3D points. For each matched pixel pip_i in I1I_1, we unproject its ground-truth depth with intrinsics K1K_1 to obtain XiX_i in the camera-1 frame.
  • 2D targets. We obtain xix_i in I2I_2 using MASt3R's fast reciprocal matching (Leroy et al., 2025, Sec. 3.3) on the dense descriptor maps d1,d2\mathbf{d}_1, \mathbf{d}_2 from the frozen MASt3R heads. A match is a pair of pixels that are mutual nearest neighbors in descriptor space.
\text{ and } p = \mathrm{NN}_1(\mathbf{d}_2(x))\bigr\\}, \qquad \mathrm{NN}_2(\mathbf{d}) = \arg\max_{x \in I_2} \langle \mathbf{d}_2(x), \mathbf{d} \rangle,$$ We define $\mathrm{NN}_1$ in the same way over the pixels of $I_1$. The search finds matches iteratively. 1. **Seeds.** Initialize $U^0$ with the pixels of a regular grid with stride $s$ in $I_1$. 2. **Forward and backward search.** Map every pixel of $U^t$ to its nearest neighbor among *all* pixels of $I_2$, giving $V^t$. Map every pixel of $V^t$ back to its nearest neighbor among all pixels of $I_1$, giving $U^{t+1}$. 3. **Cycle check.** A pixel that returns to where it started ($U^{t+1}_n = U^t_n$) forms a cycle. Keep $(U^t_n, V^t_n)$ as a reciprocal match and remove it from the active set. The other pixels continue from their new locations $U^{t+1}_n$. 4. **Termination.** Repeat steps 2–3 for a fixed number of iterations. Discard pixels that have not converged and remove duplicates, since several seeds can converge to the same match. The output $\\{(p_i, x_i)\\}$ is a subset of $\mathcal{M}$, obtained in $O(kWH)$ instead of $O(W^2H^2)$ operations for $k$ seeds. The search covers the entire image, so matches can occur at locations far apart and outside the seed grid. Matching takes only the descriptor maps as input, so it is independent of the predicted and ground-truth poses. **(2) Mapping of descriptor pairs and weights to PnP correspondences** - The “subsampled grid with stride $s$” in Section 3.5 specifies **seed pixels** for fast reciprocal matching. The search determines the corresponding pixel. - Each reciprocal match $(p_i, x_i)$ defines one 3D–2D correspondence $(X_i, x_i)$ and one FusionTransformer weight $w_i$. The weight is the RANSAC sampling probability of that correspondence. - The predicted pose has two uses. We project the source 3D points under this pose for the descriptor loss $\mathcal{L}_{descriptor}$ (Eq. 8). The RANSAC prior scores each PnP hypothesis by its Frobenius distance to the regressed pose. Figure 3 shows both paths. The PnP inputs come from the reciprocal matches and depth. - The same weights $w_i$ enter the loss in Section 3.6, $\mathcal{L}_{descriptor} = -\sum_i w_i \cdot \mathrm{sim}(\mathbf{d}_i^{src}, \mathbf{d}_i^{proj})$ (Eq. 8). This term supplies the FusionTransformer gradient. The solver output $\mathbf{P}_{\text{solver}}$ is detached. RANSAC uses $w_i$ as sampling probabilities and receives no gradients. **(3) Source of the corrective signal** During training, the FusionTransformer predicts a weight $w_i$ for each correspondence. Weighted RANSAC-PnP (Algorithm 1) samples its minimal sets with these weights. The consistency loss pulls $\mathbf{P}_{\text{regression}}$ toward the detached target $\mathbf{P}_{\text{solver}}$. The 2D targets are **independently estimated image correspondences** from the reciprocal descriptor matching in (1). The solver uses image matches and ground-truth depth, so its output can differ from $\mathbf{P}_{\text{regression}}$. The predicted pose enters the solver only through the soft prior, which breaks ties in inlier count at the default setting (Q4(3)). The consistency loss therefore supplies supervision from the image matches and ground-truth depth, avoiding circular supervision. In the offline analysis in Q3, which uses uniform sampling, $\mathbf{P}_{\text{solver}}$ has lower rotation error than $\mathbf{P}_{\text{regression}}$ on 82.6% of evaluated pairs at epoch 5. The median errors are $0.83^\circ$ and $1.06^\circ$, respectively. **Revisions.** In the revised paper (changes marked in red), Figure 3 shows that reciprocal matches supply the observations for PnP and that predicted-pose projection is used for $\mathcal{L}_{descriptor}$. We rewrote Section 3.5, stated the coordinate conventions in Algorithm 1, and added Appendix A.3 on how we construct the 3D–2D correspondences.
Reviewer_aUJ5OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

GeLoc3r adds training-time geometric regularization to a relative pose regressor while preserving its inference architecture. It combines descriptor reprojection supervision with a weighted-RANSAC pose consistency loss. Experiments report improvements over a same-data ReLoc3R baseline on relative pose estimation and visual localization.

Critical strengths

Improving geometric accuracy without adding inference-time computation is a useful goal. The same-data ReLoc3R baseline provides evidence that the overall training changes improve performance.

Critical weaknesses
  1. The central solver-based supervision appears circular. According to Figure 3, 3D points are projected using the predicted pose to generate the 2D observations supplied to PnP. These correspondences already satisfy that prediction, so solving them should recover essentially the same pose rather than provide an independent geometric correction. Detaching the solver does not resolve this issue. As written, the paper does not establish why this loss provides meaningful supervision.

  2. The comparisons do not adequately support the state-of-the-art regression claim. There is no controlled comparison with n-view end-to-end pose methods. Appendix C only quotes VGGT results with different view counts and AUC thresholds, which the authors acknowledge are not directly comparable. This is particularly relevant to localization, where GeLoc3r already retrieves ten references. Pairwise and multi-view methods should be evaluated with clearly specified inputs, common metrics, and runtime reporting.

Questions for authors

Are the PnP target pixels obtained from independent image correspondences, or generated by projecting with the predicted pose? If the latter, what prevents the consistency loss from being trivial?

How does GeLoc3r compare with feed-forward multi-view methods under matched evaluation protocols, including the ten-reference localization setting?

2 comments

We thank the reviewer for the careful reading. Both concerns are addressed below.

W1/Q1. Is the solver-based supervision circular?

No, and we thank the reviewer for raising this point. The PnP targets are independent image correspondences, not projections under the predicted pose.

  • 3D points. For each matched pixel pip_i in I1I_1, we unproject its ground-truth depth with intrinsics K1K_1 to obtain XiX_i in the camera-1 frame.

  • 2D targets. The corresponding location xix_i in I2I_2 is obtained by MASt3R's fast reciprocal matching (Leroy et al., 2025, Sec. 3.3) on the dense descriptor maps d1,d2\mathbf{d}_1, \mathbf{d}_2 from the frozen MASt3R heads. A pair (pi,xi)(p_i, x_i) is kept only if the two pixels are mutual nearest neighbors in descriptor space:

    \quad\text{and}\quad p_i = \arg\max_{p \in I_1} \langle \mathbf{d}_2(x_i), \mathbf{d}_1(p) \rangle .$$ The search starts from a regular grid of seed pixels in $I_1$, maps each seed to its nearest neighbor among all pixels of $I_2$ and back among all pixels of $I_1$, keeps the pixels that return to where they started, and iterates on the others. Its only inputs are the two descriptor maps. No camera pose, predicted or ground-truth, enters the matching, so $x_i$ does not depend on $\mathbf{P}_{\text{regression}}$.
  • Where the predicted pose is used. We project the source 3D points under the regressed pose for the descriptor loss Ldescriptor\mathcal{L}_{descriptor} (Eq. 8). The RANSAC prior scores each PnP hypothesis by its Frobenius distance to the regressed pose. Neither use generates PnP inputs.

The solver estimates Psolver\mathbf{P}_{\text{solver}} from reciprocal image matches and ground-truth depth, so it can differ from Pregression\mathbf{P}_{\text{regression}}. The consistency loss uses the solver estimate as a detached target for the regressor. We compared both poses against ground truth using a sample of 1,800 training pairs (200 from each of the nine training datasets with depth). The offline solver used the PnP and scoring settings of Algorithm 1, reciprocal matches from the released MASt3R model, and uniform sampling.

CheckpointMedian rotation error (°), regression / solverMedian translation-direction error (°), regression / solverPairs improved (%), rotation / translation
Early (epoch 5)1.06 / 0.832.15 / 1.8182.6 / 74.2
Middle (epoch 48)0.58 / 0.451.12 / 1.0470.0 / 57.7
Final (epoch 92)0.37 / 0.330.85 / 0.8554.6 / 48.9

In this offline comparison, at epoch 5 the solver had lower rotation error on 82.6% of evaluated pairs and lower translation-direction error on 74.2%. The two estimates had similar median errors by epoch 92. Q3 in our response to Reviewer bEQR reports the sizes of the improvements and degradations.

In the revised manuscript, the red text in Section 3.5 explains reciprocal descriptor matching and how each match receives a RANSAC sampling weight. In Section 3.6, the red text under “Direct Supervision” specifies the 3D–2D inputs to PnP and the role of the pose prior. Under “Indirect Supervision,” it defines the projected descriptors used in Eq. 8. The revised Figure 3 and its caption show both paths. Algorithm 1 in Appendix A.1 states the camera-frame conventions, and Appendix A.3 details the construction of the 3D–2D correspondences.

W2/Q2. Comparison with feed-forward multi-view methods. We have evaluated VGGT under the protocols of Tables 1–3, with the same image pairs and the same retrievals.

Pairwise relative pose. Both methods receive the same image pairs (512×384 for GeLoc3r; resized to width 518 for VGGT, portrait images upright, as in VGGT's preprocessing). AUC@5°/10°/20°:

MethodScanNet1500MegaDepth1500RealEstate10KTime (ms)Peak mem. (GB)
VGGT (2 views)35.34 / 58.10 / 74.9339.80 / 60.49 / 76.4329.14 / 50.48 / 68.39115†7.0
GeLoc3r-512 (Table 1)34.40 / 58.12 / 75.5650.45 / 69.29 / 82.5568.66 / 81.49 / 89.18332.6

Time per pair on an RTX 4090 (Table 1). † VGGT with matching. Peak GPU memory per pair.

Ten-reference localization. Both methods use the same NetVLAD top-10 retrievals. VGGT receives the query and its ten references jointly (11 views); the ten query-to-reference poses it predicts are passed to the same motion averaging as ours. GeLoc3r makes ten pairwise predictions followed by motion averaging. Median translation / rotation error (m/°) and peak GPU memory per query:

7-ScenesChessFireHeadsOfficePumpkinKitchenStairsAvgPeak mem. (GB)
VGGT (11 views)0.03/0.780.02/0.940.01/0.770.04/0.830.05/1.060.04/1.260.03/0.780.03/0.929.1
GeLoc3r-512 (10 pairs, Table 2)0.03/0.800.02/0.820.01/0.830.04/0.820.05/1.000.04/1.220.07/1.740.04/1.033.3
CambridgeGreatCourtKingsCollegeOldHospitalShopFacadeStMarysChurchAvgPeak mem. (GB)
VGGT (11 views)0.61/0.380.38/0.380.49/0.450.10/0.300.27/0.390.37/0.389.1
GeLoc3r-512 (10 pairs, Table 3)1.07/0.520.41/0.340.64/0.540.15/0.640.27/0.580.51/0.523.3

Cost. GeLoc3r needs 2.7–2.8 times less GPU memory than VGGT: its peak memory is 2.6 GB per pair and 3.3 GB per query (ten pairs in one batch), against 7.0 GB and 9.1 GB for VGGT. GeLoc3r is also 3.5 times faster per pair than VGGT with matching (33 ms against 115 ms on an RTX 4090, Table 1), and it has 2.6 times fewer parameters (0.43 B in the pose path, against 1.13 B in VGGT's aggregator and camera head).

In summary, with joint 11-view inference and a model 2.6 times larger, VGGT is more accurate on both localization benchmarks (7-Scenes: 0.03 m/0.92° vs. 0.04 m/1.03°; Cambridge: 0.37 m/0.38° vs. 0.51 m/0.52°), with the largest difference on Stairs, and the two are comparable on pairwise ScanNet1500. GeLoc3r is more accurate on pairwise MegaDepth1500 (+10.7 AUC@5°) and RealEstate10K, is 3.5 times faster per pair than VGGT with matching, has 2.6 times fewer parameters, and needs 2.7–2.8 times less GPU memory. We have replaced Appendix C with this comparison (changes marked in red in the revised paper).

Reviewer_ftgyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GeLoc3r introduces geometric consistency regularization in ReLoc3R, a state-of-the-art relative camera pose regression method. Different from prior work such as FAR, GeLoc3r introduces geometric consistency regularization in training process only to preserve the inference speed. Extensive experiments on seven datasets verify the effectiveness of GeLoc3r.

Critical strengths
  • In visual localization, pose regression is not robust when compared to feature matching methods. Therefore, utilizing geometric consistency in pose regression makes a lot of sense. Although FAR already introduced this hybrid approach in inference, GeLoc3r modified the training process instead to preserve the test-time efficiency, which is important for applications.

  • The evaluation is solid. GeLoc3r outperforms ReLoc3R on seven benchmarks, while keeping the same inference speed.

Critical weaknesses
  • The hybrid scoring function in Sec. 3.6 consists of N_inliers and a prior term. The number of inliers could be up to 10^2-10^3, while the prior term is at most 0.5 according to the paper. Therefore, it seems that the prior could be useless in most cases. Could the authors add an ablation study about the choice of beta and tau.

  • Could the authors clarify the design of FusionTransformer. Why descriptors at the same spatial location in both views after subsampling are treated as potential correspondences? If the pose difference of two images is huge, it is highly possible that the assumption is wrong.

  • As discussed in related works, FAR is a baseline that shares similar hybrid framework. It should be evaluated and compared on benchmarks.

Questions for authors

First row of Fig. 1: the co-visible region should be highlighted with a different color for better visual quality

2 comments

We thank the reviewer for the positive assessment and constructive suggestions.

W1. β\beta and τ\tau in the hybrid score. With the default β=0.5\beta = 0.5, the prior term can never outweigh a single inlier, so it only breaks ties in inlier count. We measured the effect of β\beta and τ\tau at the solver on 1,752 training pairs from the nine training datasets with depth, using the final checkpoint. All settings score the same 50 hypotheses per pair. Each cell gives the median rotation / translation-direction error of Psolver\mathbf{P}_{\text{solver}} against ground truth in degrees (lower is better) and, in parentheses, the percentage of pairs whose selected hypothesis differs from β=0\beta = 0:

τ=1\tau = 1τ=10\tau = 10τ=100\tau = 100
β=0\beta = 00.334 / 0.847 (–)0.334 / 0.847 (–)0.334 / 0.847 (–)
β=0.5\beta = 0.50.330 / 0.846 (11.5%)0.329 / 0.846 (11.9%) (default)0.329 / 0.846 (11.9%)
β=5\beta = 50.330 / 0.846 (11.6%)0.329 / 0.846 (12.0%)0.329 / 0.846 (12.0%)
β=50\beta = 500.333 / 0.849 (12.4%)0.330 / 0.848 (13.1%)0.329 / 0.846 (12.3%)

At the default setting, the prior changes the selected hypothesis on 11.9% of pairs relative to β=0\beta = 0. Across the tested settings, the median rotation and translation-direction errors varied by at most 0.005∘0.005^\circ and 0.003∘0.003^\circ, respectively. We have revised Section 3.6 and Appendix A.1 to explain the role of the prior at β=0.5\beta = 0.5 (changes marked in red).

W2. Design of the FusionTransformer. We thank the reviewer for raising this point. Descriptors at the same grid location are not treated as correspondences. The “subsampled grid with stride ss” refers to the seed pixels of MASt3R's fast reciprocal matching (Leroy et al., 2025). Starting from these seeds in I1I_1, matches are searched over the entire image I2I_2, and only mutual nearest neighbors are kept. Correspondences therefore remain valid under large viewpoint changes. For each reciprocal match (pi,xi)(p_i, x_i), the FusionTransformer takes the concatenated descriptors [d1(pi);d2(xi)][\mathbf{d}_1(p_i); \mathbf{d}_2(x_i)] and outputs a weight wiw_i, which is used as the RANSAC sampling probability. We have rewritten Section 3.5 to state this explicitly and added Appendix A.3 on the construction of the correspondences (changes marked in red in the revised paper).

W3. Comparison with FAR. ReLoc3R has already evaluated FAR on the benchmarks we use. We follow the evaluation protocol of ReLoc3R on ScanNet1500 and RealEstate10K, and our pipeline reproduces the published results of the official ReLoc3R-512 checkpoint on both, so the FAR results reported there are directly comparable with ours. Dong et al. (2025, Appendix D) report FAR AUC@20$^\circ$ scores of 28.19 on ScanNet1500 and 37.67 on RealEstate10K. The corresponding GeLoc3r scores in Table 1 are 75.56 and 89.18. FAR combines a learned pose estimate with a geometric solver at inference. GeLoc3r uses geometric supervision during training and directly regresses poses at inference.

Q1. Figure 1. We have redrawn the first row of Figure 1: the co-visible region is now highlighted in a distinct color with an outline.

Thanks for the detailed response. I maintain my positive rating.

Comments on the paper

1 comment

Dear Reviewer, As the discussion period draws to a close, we would welcome your feedback on whether our response and additional results address your concerns. If any questions remain, we would be happy to clarify. Thank you for your time and thoughtful review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.