acceptodds
Under review as a conference paper at ICLR 2027

UniPS: Unified Patch Similarity Unlocks Training-Free Visual Place Recognition

Abstract

Visual Place Recognition (VPR) localises an image by retrieving the geo- referenced database image that shows the same place. Most VPR methods are training-based, aiming to learn effective features from annotated data. These methods achieve strong results on standard VPR benchmarks but require millions of labelled samples, making annotation cost a significant challenge. Even with such extensive training, they still fail on other VPR benchmarks that differ from their training set, such as cross-environment benchmarks. Training-free VPR methods avoid this annotation cost and generalise better. However, these methods do not fully leverage pretrained vision models, resulting in features that lack sufficient discriminative power to achieve strong performance on standard VPR benchmarks, highlighting the challenge of effectively utilising pretrained vision models. To tackle these challenges, we propose UniPS (Unified Patch Similarity), a training-free pipeline built solely on patch similarity. Given a query or database image, only a frozen backbone is applied to extract patch features. We then perform a series of patch similarity-based comparisons for VPR: self-similarity for patch importance, followed by weighted second-moment patch feature pooling for coarse ranking, and cross-image patch similarity for reranking. Based on comprehensive evaluation on 17 benchmarks, UniPS achieves state-of-the-art results on both the 8 standard and 9 cross-environment VPR benchmarks. On standard VPR benchmarks, UniPS outperforms the best training-free baseline by +12.9% and the best trained baseline by +0.2%, despite the latter’s heavy annotation requirements. On cross-environment benchmarks, UniPS outperforms the best training-free baseline by +7.1% and the best trained baseline by +11.5%. In addition, the proposed module can be easily plugged into existing methods, improving their performance by +5.6% to up to +17.8%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visual Place Recognition, Training-free, Image Retrieval

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_hSsNOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces UniPS, a training-free Visual Place Recognition (VPR) framework based on unified patch similarity. The method combines self-similarity-based patch weighting, weighted second-moment pooling for global retrieval, and patch-level similarity for reranking. The authors evaluate the approach on 17 standard and cross-environment benchmarks, reporting impressive results that outperform many trained and training-free baselines.

The extensive experimental evaluation and strong performance across diverse environments are clear strengths of the paper. However, my main concern is the limited methodological novelty, as the individual components are largely established techniques. Furthermore, several experimental comparisons require better control to establish the contribution of the proposed method.

Critical strengths
  1. The proposed training-free pipeline achieves impressive results across a broad range of VPR benchmarks, particularly in cross-environment settings.
  2. The unified framework is relatively simple and does not require additional training or labeled data.
  3. The extensive experiments, component ablations, and demonstrated applicability of the reranking module to existing methods provide useful empirical evidence.
Critical weaknesses

My main concern is that the proposed components, including self-similarity weighting, second-order pooling, and patch-based reranking, are already established techniques. The contribution therefore appears to lie primarily in their particular combination.

Moreover, the reranking formulation closely resembles EffoVPR (ICLR 2025), replacing attention-based patch selection with self-similarity weighting and MNN counting with weighted MaxSim, while even adopting the same threshold of θ=0.65\theta = 0.65. A more direct comparison would help clarify both the methodological novelty and the contribution of these modifications to performance.

Beyond the limited methodological novelty, it remains unclear how much of the reported performance improvement can be attributed to specific design choices rather than other factors. More controlled comparisons with closely related methods would help establish both the novelty and the actual contribution of the proposed components to the performance gains.

Questions for authors

Input resolution. Most baselines use input resolutions of 224 to 384 pixels, while UniPS uses 672 to 768 pixels. Beyond increasing runtime, higher resolution may also contribute to the performance gains. Please provide results at comparable resolutions.

Coarse ranking. The standalone performance of the proposed PCA-based global descriptor should be compared against CLS-token retrieval under matched settings and reported in Table 2. Reporting performance before reranking would help isolate the contribution of the proposed representation.

Feature extraction details. Appendix A.2 mentions value-facet features but does not specify the extraction layer. This is important since AnyLoc and EffoVPR-ZS use intermediate-layer features, whereas other models (e.g. SALAD, CricaVPR and more) use output-layer features. Please clarify the selected layer.

Missing top-K specification and ablation. The paper does not specify the value of K used for reranking, despite its impact on both retrieval accuracy and computational cost. Please clarify the value used and provide an ablation over different K values, reporting R@1 and runtime. Coarse-stage R@K should also be reported to distinguish candidate retrieval quality from reranking effectiveness.

Reranking generalisation. EffoVPR-ZS, the only other two-stage training-free baseline, is absent from this analysis. Given the similarity between the reranking mechanisms, applying the proposed reranker to the top-K candidates retrieved by EffoVPR-ZS would provide stronger evidence of its independent contribution.

Missing SoTA comparisons. Several relevant recent methods are absent from the quantitative comparisons, including EffoVPR, Pair-VPR (RA-L 2025), ImAge (NeurIPS 2025), SAGE (ICLR 2026), and SelaVPR++ (TPAMI 2026). Some report stronger results on individual standard VPR benchmarks. Please extend the comparison to these methods.

Minor Comments

  • Line 192: CosPlace and EigenPlaces are not trained on the entire 41M-image SF-XL dataset. Please clarify the actual training set sizes.
  • Line 202: What is the motivation or origin of the similarity template? Are the patch features L2-normalized so that their dot products correspond to cosine similarity?
  • Figure 1: The bottom-right panel could benefit from visual refinement.
  • Line 799: Missing period between sentences.
  • Line 1125: "FUTUREWORK" should be "FUTURE WORK".
12 comments
AuthorsOpenReview ↗

Response to Reviewer hSsN (Part 1/5)

Thank you for your comments. We will address them as follows:

  1. It remains unclear how much of the reported performance improvement [...] Most baselines use input resolutions of 224 to 384 pixels, while UniPS uses 672 to 768 pixels. Please provide results at comparable resolutions. (Q1 & Q4, W3)

Similar to SegVLAD, our motivation for adopting high-resolution images is to fully utilise the visual information inherently available in the original images. Many of the commonly used datasets (e.g., Pitts, MSLS-val, St Lucia, Hawkins, Laurel Caverns, 17 Places, etc.) contain images with resolutions higher than 322 × 322, whereas many training-based baselines operate at substantially lower resolutions. For example, existing trained methods typically resize images to 224 × 224 during training (e.g., SALAD) to reduce the training time, and increase the resolution to 322 × 322 during inference (e.g., SALAD and BoQ). This creates a potential information bottleneck: visual details present in the original high-resolution images may already have been irreversibly lost during downsampling, potentially leading to retrieval failures.

Appendix C.3 (Figure 12) gives an example on Pitts30k (for the same query and its correct database image, the 640×640 input (40×40 patches) yields 115 patch matches and a reranking score of 12.8; at 256×256 (16×16 patches), 95 of these 115 matches disappear, the number of matches drops to 27 and the score falls to 8.5, below the 11.0 of an image from a wrong place, which is therefore ranked first. The lost matches lie mostly on fine details such as windows on the building facade, signs and traffic lights, which can no longer be resolved at low resolution). In contrast, our training-free approach is not constrained by the long training cost and can directly make full use of the visual information available at higher resolutions, avoiding the information loss from downsampling.

In addition, we conducted a systematic comparison across different input resolutions: all methods use the same DINOv3 backbone, and only the test-time input resolution changes (322 for all methods in Table R1, 768 in Table R2). As shown in Tables R1 and R2, our method achieves state-of-the-art performance at both resolutions. More specifically, at 322, OursD3 reaches a mean R@1 of 82.7 over all 17 benchmarks, 5.4 points above the best trained method, BoQ (77.3), and 11.7 points above it on the cross-environment benchmarks (75.2 vs. 63.5). At 768, OursD3 leads on both groups: +1.1 on the standard benchmarks (93.2 vs. 92.1), +10.0 on the cross-environment benchmarks (76.6 vs. 66.6) and +5.8 over all 17 (84.4 vs. 78.6).

Table R1. All methods on the same DINOv3 backbone at 322×322 input (R@1). † marks rows reported in Table 6.

MethodDimStd (8)Cross (9)All (17)
SALAD†844890.761.875.4
BoQ†1228892.963.577.3
CliqueMining†844891.956.973.4
OursD3102491.275.282.7

Table R2. All methods on the same DINOv3 backbone at 768×768 input (R@1). † marks rows reported in Table 6.

MethodDimStd (8)Cross (9)All (17)
SALAD844889.065.176.4
BoQ1228892.166.678.6
CliqueMining844891.358.573.9
OursD3†102493.276.684.4

The full tables and analyses are included in Appendix B.4 (Tables 20 and 21).

AuthorsOpenReview ↗

Response to Reviewer hSsN (Part 2/5)

Ablation study on input resolution.

We also evaluated our method alone at different input resolutions (Table R3, with patch grids aligned between D2 and D3). Even at 322, OursD2 (79.3) and OursD3 (82.7) are above every trained method in Table R1 over the 17 benchmarks (the best is BoQ, 77.3). R@1 rises with resolution, peaking at about 672 for D2 and 640–768 for D3. Table R4 shows that at every resolution, K = 10 already captures most of the gain from reranking.

Table R3. R@1 vs. input resolution (grid-aligned, same tokens for D2 and D3).

Model @ resTokensStd (8)Cross (9)All (17)
OursD2 @22425684.370.176.8
OursD2 @32252987.871.779.3
OursD2 @33657688.274.280.8
OursD2 @448102489.275.381.8
OursD2 @560160090.175.982.6
OursD2 @672 (Table 2)230490.376.382.9
OursD2 @784313690.376.082.7
OursD3 @25625690.273.881.5
OursD3 @32240091.275.282.7
OursD3 @38457691.876.283.5
OursD3 @512102492.076.383.7
OursD3 @640160092.777.184.4
OursD3 @768 (Table 2)230493.276.684.4
OursD3 @896313692.776.384.0

Table R4. Mean R@1 against shortlist size K at each input resolution; K=1 is the coarse ranking alone.

Model @ resStd K=1Std K=10Std K=1000Cross K=1Cross K=10Cross K=1000
OursD2 @22476.483.684.368.071.570.1
OursD2 @32276.485.487.863.372.171.7
OursD2 @33676.285.788.264.074.574.2
OursD2 @44877.486.889.265.273.675.3
OursD2 @56076.687.190.165.975.875.9
OursD2 @672 (Table 2)81.287.490.365.975.276.3
OursD3 @25682.789.890.269.474.373.8
OursD3 @32283.389.991.267.675.475.2
OursD3 @38483.590.491.870.976.176.2
OursD3 @51282.190.292.070.876.676.3
OursD3 @64082.891.092.767.475.977.1
OursD3 @768 (Table 2)82.591.593.270.275.676.6
OursD3 @89680.191.092.766.175.976.3

The full tables and analyses are included in Appendix B.4 (Tables 22 and 23).

AuthorsOpenReview ↗

Response to Reviewer hSsN (Part 3/5)

  1. The paper does not specify the value of K used for reranking. Please clarify the value used and provide an ablation over different K values, reporting R@1 and runtime. Coarse-stage R@K should also be reported to distinguish candidate retrieval quality from reranking effectiveness. (Q1 & Q4, W3)

All results of our method in the paper are reported at K = 1000, although K can be adjusted according to practical deployment requirements. We further conduct an ablation study on K, reporting coarse-ranking performance (R@K), reranked R@1, and the runtime of each stage using two backbones. As shown in Tables R5 and R6, K = 10 already achieves strong performance with very low coarse-ranking and reranking costs, demonstrating the effectiveness of coarse ranking alone. Specifically, for OursD3, coarse R@10 reaches 95.5 on standard and 93.2 on cross-environment benchmarks. Reranking only 10 candidates achieves a mean R@1 of 83.1 across the 17 benchmarks, just 1.3 points below K = 1000 (84.4), while reducing the total query time from 258.8 ms to 63.9 ms (Table R7). Increasing K to 100 further raises R@1 to 84.2, within 0.2 points of K = 1000. OursD2 exhibits the same trend.

Table R5. OursD3. Coarse = coarse-ranking R@K; R@1 = after reranking; Std (8), Cross (9), All (17).

KCoarse StdCoarse CrossR@1 StdR@1 CrossR@1 All
182.570.282.570.276.0
1095.593.291.575.683.1
5098.499.292.776.684.2
10098.999.592.876.684.2
20099.499.793.176.684.4
50099.799.893.276.684.4
100099.999.993.276.684.4

Table R6. OursD2. Same columns as Table R5.

KCoarse StdCoarse CrossR@1 StdR@1 CrossR@1 All
181.265.981.265.973.1
1092.291.387.475.280.9
5096.198.389.176.582.4
10097.398.889.576.382.5
20098.299.389.976.382.7
50099.399.690.276.382.8
100099.799.890.376.382.9

Table R7. Time (ms/query) by stage, mean over the 17 benchmarks (one H100).

MethodKExtractionCoarseRerankTotal
OursD2114.30.08–14.4
1014.40.103.618.1
5014.40.1020.334.8
10014.40.1140.154.7
100014.40.12195.7210.2
OursD3160.90.10–61.0
1060.10.103.663.9
5059.50.1119.679.3
10059.50.1541.5101.3
100058.70.15199.9258.8

In addition, we added a per-query, stage-wise timing comparison of UniPS-D2/D3 and other two-stage methods across all 17 benchmarks on one H100 (Table R8). Despite using a large backbone and high resolution, our method introduces little overhead beyond feature extraction. At K = 10, reranking takes only 3.6 ms, 6.6× faster than the fastest competing reranker (EffoVPR-ZS, 23.9 ms). OursD2 achieves the lowest total latency (18.1 ms) among all two-stage methods, while outperforming the best competitor, EffoVPR-ZS, by 6.8 R@1 points (80.9 vs. 74.1). Per-stage timings are provided in Appendix A.6 (Table 12).

Table R8. Per-query cost (ms), mean over the 17 benchmarks; Performance is the mean R@1 over the 17 benchmarks.

MethodExtractionCoarseRerankingTotalPerformance
Patch-NetVLAD14.70.095189520460.7
R2Former3.90.0563.567.462.6
SelaVPR11.80.0636.648.571.1
EffoVPR-ZS6.20.0623.930.274.1
OursD2, K=1014.40.103.618.180.9
OursD3, K=1060.10.103.663.983.1
OursD2, K=10014.40.1140.154.782.5
OursD3, K=10059.50.1541.5101.384.2
OursD2, K=100014.40.12195.7210.282.9
OursD3, K=100058.70.15199.9258.884.4
Reviewer_qWtrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces UniPS, a training-free framework for Visual Place Recognition (VPR). The framework leverages the self-similarity of image patches for weighting, employs weighted second-moment pooling for coarse ranking, and utilizes weighted MaxSim for reranking. The method achieves excellent performance across 17 standard and cross-environment benchmarks. Furthermore, the proposed re-ranking module can enhance the performance of existing VPR methods.

Critical strengths
  1. The research problem is valuable, the proposed method is well-motivated, and the paper is generally clearly written.
  2. The experimental evaluation is comprehensive, covering 17 diverse benchmarks. Additionally, Table 4 shows that the proposed reranking module improves the average performance of 10 existing VPR methods.
Critical weaknesses
  1. The fairness and clarity of the experimental comparisons are insufficient. UniPS-D2 and UniPS-D3 use input resolutions of 672×672 and 768×768, respectively, whereas baseline methods such as SALAD and BoQ typically use 322×322. Moreover, the paper does not explicitly specify which variants of the DINOv2/DINOv3 models were used. Consequently, it is difficult to determine to what extent the performance gains are attributable to the proposed method itself.
  2. The analysis of efficiency and storage cost is incomplete. Table 7 only reports the inference time for UniPS-D2, while the computational cost of the best-performing variant, UniPS-D3, remains unclear. Also, the 1024-dimensional global descriptor reported in the paper does not account for the local patch features required for the reranking stage. Therefore, the actual database storage cost of the method remains ambiguous.
Questions for authors
  1. Could you clarify which specific variants of the DINOv2/DINOv3 models were used in this work? Could you provide a comparative experiment with controlled input resolution (e.g., evaluating UniPS-D2 at a resolution of 322×322)? Additionally, I noticed that on the Nordland dataset, the R@1 for UniPS-D3 shows a significant improvement over UniPS-D2, i.e., increasing from 82.3% to 96.0%. What is the reason for this substantial gain?
  2. What hardware and top-K setting are used? What data are used to fit PCA and whitened PCA, and are these transformations shared across all 17 benchmarks?
  3. Could you provide the inference time for each stage of UniPS-D3? Furthermore, could you specify the database storage requirements for the proposed method, including the storage cost for the local patch features needed for reranking?
  4. Table 4 only reports results for the re-ranking module on cross-environment benchmarks. Could you provide additional experimental results on standard VPR benchmarks to verify whether the module maintains its performance on those datasets as well?
  5. Section 4.2 states that the AnyLoc's descriptor is 24× larger than UniPS's, but Table 2 lists the dimensions as 49152 and 1024, respectively, corresponding to 48×. Please verify this inconsistency.
14 comments
AuthorsOpenReview ↗

Response to Reviewer qWtr (Part 1/7)

Thank you for your questions. We will address your comments as follows:

  1. UniPS-D2/D3 use 672×672 and 768×768 inputs, while baselines such as SALAD and BoQ use 322×322. (W1 & Q1)

Similar to SegVLAD, our motivation for adopting high-resolution images is to fully exploit the visual information available in the original images. Many commonly used datasets (e.g., Pitts, MSLS-val, St Lucia, Hawkins, Laurel Caverns, and 17 Places) contain images at resolutions higher than 322 × 322, while many training-based baselines use substantially lower resolutions. Such downsampling may irreversibly discard useful visual details, potentially creating an information bottleneck and leading to retrieval failures.

Appendix C.3 (Figure 12) provides an example on Pitts30k: for the same query and correct database image, increasing the input resolution from 256×256 (16×16 patches) to 640×640 (40×40 patches) increases the matches from 27 to 115 and the reranking score from 8.5 to 12.8, surpassing the incorrect image's score of 11.0. At low resolution, 95 of the 115 matches are lost, primarily due to fine details such as windows and signs becoming unresolvable. In contrast, our training-free approach avoids costly training and can directly exploit high-resolution visual information without the information loss caused by downsampling.

In addition, we compare different input resolutions, keeping the backbone fixed and varying only the test-time resolution (322 × 322 in Table R1 and 768 × 768 in Table R2). Our method achieves state-of-the-art performance at both resolutions. At 322 × 322, it outperforms the best trained method, BoQ, by 5.4 points across all 17 benchmarks (82.7 vs. 77.3) and 11.7 points on cross-environment benchmarks (75.2 vs. 63.5). At 768 × 768, it leads by 1.1, 10.0, and 5.8 points on standard (93.2 vs. 92.1), cross-environment (76.6 vs. 66.6), and all benchmarks (84.4 vs. 78.6), respectively.

Table R1. All methods on the same DINOv3-7B backbone at 322×322 input (R@1). † marks rows reported in Table 6.

MethodDimStd (8)Cross (9)All (17)
SALAD†844890.761.875.4
BoQ†1228892.963.577.3
CliqueMining†844891.956.973.4
OursD3102491.275.282.7

Table R2. All methods on the same DINOv3-7B backbone at 768×768 input (R@1). † marks rows reported in Table 6.

MethodDimStd (8)Cross (9)All (17)
SALAD844889.065.176.4
BoQ1228892.166.678.6
CliqueMining844891.358.573.9
OursD3†102493.276.684.4

The full tables and analyses are included in Appendix B.4 (Tables 20 and 21). Full image resolution ablation study of the proposed method (from 224×224 to 896×896) is also included in Appendix B.4 (Table 22).

AuthorsOpenReview ↗

Response to Reviewer qWtr (Part 2/7)

  1. Which DINOv2/DINOv3 version is used is not stated. (W1 & Q1)

In this paper, UniPS-D2 uses DINOv2-G, following previous training-free methods such as AnyLoc and SegVLAD. As our method keeps the backbone frozen, upgrading to a stronger foundation model requires no retraining. We therefore introduce UniPS-D3 with DINOv3-7B and evaluate DINOv3-L and DINOv3-7B to assess whether stronger foundation models translate into better VPR performance. As shown in Table R4, upgrading from DINOv2-B to DINOv3-7B improves our mean R@1 across 17 benchmarks by 9.1 points, compared with only 1.4–3.6 points for trained methods using the same backbones.

Compared with trained methods using the same backbone (Table R3), our method performs particularly well with larger backbones: it outperforms AnyLoc by 12.1 points with DINOv2-G (82.9 vs. 70.8), BoQ by 7.1 points with DINOv3-7B (84.4 vs. 77.3), and SAGE by 1.5 points with DINOv2-L (77.0 vs. 75.5), achieving the best mean R@1 across all 17 benchmarks in each case. With the smaller DINOv2-B, MegaLoc and BoQ achieve higher overall means (77.6 and 75.8 vs. 75.3), while our cross-environment mean (64.6) remains within 0.5 points of MegaLoc (65.1), despite the latter being trained on data from a wide range of datasets.

Table R3. Same backbone (R@1): trained methods against ours on the same backbone, grouped by backbone. † two-stage; ‡ inductive.

BackboneMethodDimStd (8)Cross (9)All (17)
ViT-B (DINOv2-B)SALAD844889.655.971.8
CricaVPR‡1075286.753.369.0
BoQ1228891.961.475.8
CliqueMining844891.354.972.0
SuperVLAD‡307288.351.168.6
VLAD-BuFF409690.357.372.8
MegaLoc844891.765.177.6
Ours, DINOv2-B†102487.364.675.3
ViT-L (DINOv2-L)SelaVPR†102488.555.871.1
SAGE‡844893.060.075.5
SelaVPR++†512b + 409692.758.974.8
EffoVPR-ZS†102480.368.774.1
Ours, DINOv2-L†102486.268.877.0
ViT-G (DINOv2-G)AnyLoc4915272.169.570.8
SegVLAD-PreT4915275.455.264.7
Ours, DINOv2-G†102490.376.382.9
ViT-L (DINOv3-L)Ours, DINOv3-L†102491.475.482.9
ViT-7B (DINOv3-7B)SALAD844890.761.875.4
BoQ1228892.963.577.3
CliqueMining844891.956.973.4
Ours, DINOv3-7B†102493.276.684.4

The full tables and analyses are included in Appendix B.5 (Table 24).

AuthorsOpenReview ↗

Response to Reviewer qWtr (Part 3/7)

  1. It is difficult to determine to what extent the performance gains are attributable to the proposed method itself. (W1 & Q1)

The proposed training-free method aims to fully utilise foundation models through patch similarity, benefiting from better patch consistency, larger backbones and higher resolution. Nevertheless, it still outperforms most trained methods even with weaker patch consistency, smaller backbones and lower resolution (Tables R1–R3).

In addition, our method better "unlocks" the potential of stronger backbones. As shown in Table R4, scaling to ViT-7B (DINOv3-7B) improves our mean R@1 across 17 benchmarks by 9.1 (+12.0 on cross-environment benchmarks), compared with only 1.4–3.6 for the three trained methods. This demonstrates that our advantage increases with backbone scale.

Table R4. Scaling the backbone from ViT-B (DINOv2-B) to ViT-7B (DINOv3-7B), from Table R3. All = mean R@1 over the 17 benchmarks; Δ = ViT-7B − ViT-B on Std (8), Cross (9) and All (17). † two-stage.

MethodViT-B AllViT-7B AllΔ StdΔ CrossΔ All
SALAD71.875.4+1.1+5.9+3.6
BoQ75.877.3+1.0+2.1+1.5
CliqueMining72.073.4+0.6+2.0+1.4
Ours†75.384.4+5.9+12.0+9.1

The full tables and analyses are included in Appendix B.5 (Table 25).

  1. Performance on the Nordland dataset. What is the reason for this substantial gain? (Q1)

Nordland is one of the most challenging VPR benchmarks, as shown in Table 2 and 3 in the main submission. All training-free methods achieve below 58 R@1, substantially lower than on other benchmarks. Trained methods also achieve relatively low performance, with recent approaches requiring intensive training to obtain comparable results. We attribute this difficulty to seasonal variations and the lack of distinctive matching regions, such as building facades and traffic signs.

Therefore, effective patch similarity modelling is crucial for Nordland. This motivates our method, which significantly outperforms other training-free approaches (82.3 vs. 57.9 R@1). More importantly, DINOv3 offers improved patch consistency over DINOv2 through Gram anchoring, which preserves the similarity structure of patch features (Siméoni et al., 2025). Combined with our method, it further addresses the patch-matching challenges of Nordland. With the same ViT-L backbone size, replacing DINOv2-L with DINOv3-L raises our Nordland R@1 from 63.4 to 87.6 (+24.2), and the larger DINOv3-7B further boosts it to 96.

References

Oriane Siméoni et al. DINOv3. arXiv preprint arXiv:2508.10104, 2025.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Reviewer uC71 (Part 1/2)

Thank you for your questions. We will address your comments as follows.

  1. What exact image set fits the 512-dimensional patch PCA and the 1,024-dimensional whitened PCA in every benchmark? Are the transformations frozen across all 17 benchmarks, and are any evaluated query images used in fitting? (W1 & Q1)

We follow AnyLoc to fit PCA and whitened PCA on unlabelled deployment-database images, without using labels or evaluated query/test images. The same unsupervised procedure is applied to all benchmarks. The fitting data are listed in Table R1.

Table R1. How the two PCAs are fitted (token PCA to 512 channels, descriptor WPCA to 1,024-d)

Token PCA (512)Descr. WPCA (1024)
Fitting dataPatch tokens of at most 400 database imagesDescriptors of at most 6,000 database images
MeanOver the 400 sampled imagesOver the whole database

Since our benchmarks cover diverse environments (urban, indoor, aerial, subterranean, underwater, etc.), we initially fitted each benchmark separately. To test the impact of this choice, we added an external-fit control: both PCAs are fitted only once on SF-XL val (used to calibrate τ and θ) and then frozen across all 17 benchmarks, with all other settings unchanged. As shown in Table R2, both schemes yield nearly identical results: the shared fit changes the mean R@1 by only −0.2 (D3) and +0.1 (D2), with the largest drop on Nordland (−2.7 and −1.9, respectively). The cross-environment mean remains unchanged or slightly improves (+0.0 for D3 and +0.5 for D2). OursD3 (84.2) also retains a 6.9-point lead over the best trained method using the same backbone, BoQ (77.3).

Table R2. Per-benchmark fit vs. one shared external fit (R@1).

MethodPCA fitNordlandStd (8)Cross (9)All (17)
OursD3†Per-benchmark96.093.276.684.4
OursD3†Shared (SF-XL)93.392.876.684.2
Change−2.7−0.40.0−0.2
OursD2†Per-benchmark82.390.376.382.9
OursD2†Shared (SF-XL)80.490.176.883.0
Change−1.9−0.2+0.5+0.1

The full tables and analyses are included in Appendix A.4 (Tables 7 and 8).

  1. What is K, and is it fixed across all datasets (W1 & Q1)

K is fixed to 1000 on all benchmarks and is not tuned per dataset. We additionally ran K = 10 and K = 100; the results are given in Tables R3–R5.

Table R3. OursD3. Coarse = coarse-ranking R@K; R@1 = after reranking; Std (8), Cross (9), All (17).

KCoarse StdCoarse CrossR@1 StdR@1 CrossR@1 All
1095.593.291.575.683.1
10098.999.592.876.684.2
100099.999.993.276.684.4

Table R4. OursD2. Same columns as Table R3.

KCoarse StdCoarse CrossR@1 StdR@1 CrossR@1 All
1092.291.387.475.280.9
10097.398.889.576.382.5
100099.799.890.376.382.9

Table R5. Time (ms/query) by stage, mean over the 17 benchmarks (one H100).

MethodKExtractionCoarseRerankTotal
OursD21014.40.103.618.1
10014.40.1140.154.7
100014.40.12195.7210.2
OursD31060.10.103.663.9
10059.50.1541.5101.3
100058.70.15199.9258.8

The full tables and analyses are included in Appendix A.5 (Tables 9–11).

AuthorsOpenReview ↗

Response to Reviewer uC71 (Part 2/2)

  1. In Table 4, does the added reranker reuse each baseline's own patch features, or introduce UniPS's frozen DINO features and patch weights? (W2 & Q2)

In Table 4, the added reranker extracts patch features using each baseline's own backbone. For a fair comparison, all baselines use a unified resolution of 504 px for patch feature extraction, requiring an additional forward pass. The corresponding per-query extraction and reranking times are reported in Tables R6–R7.

Table R6. Extra per-query cost (ms, one H100) of the proposed reranking module at 504 px; the block used is given in brackets. Performance = mean R@1 over the 9 cross-environment benchmarks.

MethodBackboneRes.Extra extr.Rerank K=100Performance
SelaVPRDINOv2-L + adapters (22/24)50418.48.965.5
SALADDINOv2-B (10/12)5043.36.870.8
CricaVPRDINOv2-B (10/12)50410.16.867.7
BoQDINOv2-B (10/12)5043.36.872.5
CliqueMiningDINOv2-B (10/12)5043.36.870.5
SuperVLADDINOv2-B (10/12)5046.66.868.9
MegaLocDINOv2-B (10/12)5043.36.974.2
SAGEDINOv2-L (22/24)50416.38.970.7
VLAD-BuFFDINOv2-B (10/12)5043.36.869.8
EffoVPR-ZSDINOv2-L (22/24)5046.08.971.2

In practical deployment, our reranker reuses the patch features computed by each baseline's first stage, requiring only weight computation and adding negligible inference overhead. As shown in Table R7, weight computation takes under 0.5 ms, while reranking 100 candidates takes 0.5–2.1 ms (8.9 ms for EffoVPR-ZS, which already operates at 504 px). For the single-stage baselines (SALAD, BoQ, CliqueMining, MegaLoc and VLAD-BuFF), our reranker improves cross-environment mean R@1 by 6.5–12.8 points. Both cost tables are included in Appendix B.7 (Tables 30 and 31).

Table R7. Extra per-query cost (ms, one H100) at each baseline's own resolution, reusing its patch features. Performance = mean R@1 over the 9 cross-environment benchmarks.

MethodRes.Extra (w only)Rerank K=100Performance
SelaVPR2240.230.665.6
SALAD3220.431.767.6
CricaVPR2240.290.566.6
BoQ3220.301.769.2
CliqueMining3220.401.767.7
SuperVLAD3220.271.767.6
MegaLoc3220.301.771.6
SAGE3220.292.171.4
VLAD-BuFF3220.331.766.1
EffoVPR-ZS5040.188.971.2
  1. For at least one representative baseline, can the authors compare the proposed weighting/gated-MaxSim rule with a standard local matching rule using exactly the same tokens and shortlist? (W2 & Q2)

We compare our weighted, gated MaxSim rule against three standard local matching rules (MNN, MNN with EffoVPR’s 0.65 threshold, and Patch-NetVLAD’s official MNN + RANSAC) on three representative baselines, using identical tokens and candidate lists (Table R8). Our rule consistently performs best, outperforming the strongest standard rule (MNN > 0.65) by 2.1–6.6 points, while RANSAC is least effective on these cross-environment benchmarks. This comparison is included in Appendix B.7 (Table 32).

Table R8. Scoring rules on the same tokens and shortlist; mean R@1 over 9 cross-environment benchmarks. MNN = mutual nearest neighbours (SelaVPR); MNN>0.65 = EffoVPR's similarity threshold; MNN + RANSAC = Patch-NetVLAD official matcher.

BaselineFirst stageOursMNNMNN>0.65MNN + RANSAC
SALAD55.970.864.064.257.7
MegaLoc65.174.263.770.760.3
EffoVPR-ZS51.671.266.469.160.8

Dear Reviewer uC71,

Thank you again for taking the time to review our paper and share your comments. We hope our earlier response has addressed your concerns. Since the ICLR discussion period ends soon, we would greatly appreciate your feedback on our response, and we are happy to answer any further questions.

Best regards,

The Authors

Dear Chairs and Reviewers,

We sincerely thank you for the thoughtful and constructive feedback. Here is a summary for your convenience.

Overall Outcome

All three reviewers recognised the strong empirical results of UniPS. Two reviewers rate the paper as weak acceptance (hSsN, uC71), and Reviewer qWtr stated: "I will not oppose acceptance of this paper in the AC-reviewers discussion if other reviewers accept." We addressed all concerns with additional experiments and incorporated the updates in the revised manuscript.

Strengths Highlighted by Reviewers

  1. The research problem is valuable, the method is well-motivated, and the paper is generally clearly written (qWtr).
  2. UniPS is simple (hSsN, uC71) and requires no additional training or labelled data (hSsN).
  3. All three reviewers noted the comprehensive evaluation on 17 diverse benchmarks, which uC71 described as "a consequential practical contribution".
  4. All three reviewers recognised the strong results, particularly in cross-environment settings (hSsN, uC71), and the benefit of the reranking module when plugged into existing methods (qWtr, hSsN).
  5. hSsN noted that at K=10, UniPS "still achieves the highest overall mean R@1 and cross-environment mean, while remaining competitive on standard benchmarks", and that coarse ranking alone already outperforms all reported baselines on the cross-environment benchmarks.
  6. hSsN and uC71 highlighted the component ablations, and uC71 noted that the same-backbone comparison with retrained baselines "supports the value of the integration".

Weaknesses Raised and Our Revisions

Top-K as the main evaluation setting.

  • To address all three reviewers' questions about K, we added an ablation over K and, following Reviewer hSsN, now use K=10 for all main results, conclusions and generalisation experiments, with K=1000 kept as supplementary analysis. UniPS still achieves the highest mean R@1 over all 17 benchmarks, outperforms the best training-free baseline by +11.2% on the standard benchmarks while remaining competitive with trained methods, and achieves state-of-the-art cross-environment results (+6.1% and +10.5%).
  • To address Reviewer qWtr's concern about efficiency, we added per-stage runtime and database storage. Reranking takes only 3.6 ms per query, 6.6× faster than EffoVPR-ZS.
  • Please check Responses to Reviewers hSsN and qWtr, and revised Sec. 4.1 and App. A.5, A.6, B.9.

Input resolution.

  • To address Reviewers qWtr's and hSsN's concern about input resolution, we evaluated UniPS, SALAD, BoQ and CliqueMining with the same backbone and descriptor dimension, at each input resolution from 322×322 to 768×768. UniPS achieves the highest mean R@1 over all 17 benchmarks at every resolution and surpasses them on the standard benchmarks from 512×512 upward.
  • To address Reviewer qWtr's question on why trained methods drop at higher resolution, we showed that this is caused by single-resolution fine-tuning, supported by a feature-drift analysis and a SALAD retrained at 512×512.
  • Please check Responses to Reviewers qWtr and hSsN, and revised App. B.4.

Backbone.

  • To address Reviewer qWtr's question about the backbone, we clarified the variants and compared UniPS with trained methods on the same backbone. With the same backbone, UniPS achieves the highest mean R@1 over all 17 benchmarks from ViT-L upward.
  • Please check Response to Reviewer qWtr and revised App. B.5.

PCA fitting protocol.

  • To address Reviewers qWtr's and uC71's concern about PCA fitting, we clarified that both PCAs use unlabelled database images only; one shared external fit for all 17 benchmarks gives nearly identical results.
  • Please check Responses to Reviewers qWtr and uC71, and revised App. A.4.

Novelty and attribution of the components.

  • To address Reviewer hSsN's concern about the similarity to EffoVPR, we showed that our coarse descriptor clearly outperforms the [CLS] token, and that our reranker improves EffoVPR-ZS on its own shortlist and features.
  • To address Reviewer uC71's concern about attribution, we showed that our matching rule outperforms MNN and RANSAC on the same tokens and shortlist.
  • Please check Responses to Reviewers hSsN and uC71, and revised App. B.6 and B.7.

Additional analyses and results.

  • Comparison with Pair-VPR, ImAge, SAGE and SelaVPR++ (Response to Reviewer hSsN; App. B.8)
  • Reranker on the standard benchmarks (Response to Reviewer qWtr; App. B.7)
  • Nordland gain from DINOv2 to DINOv3 (Response to Reviewer qWtr)
  • Feature layer, training-set sizes and similarity template (Response to Reviewer hSsN)

Finally, we thank the area chairs and reviewers once again for their time and constructive suggestions. We hope our work contributes to more practical and generalisable visual place recognition.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.