MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering

Response to Q1:

Yes. CosPlace is our most direct baseline, and our work can be seen as replacing its fixed, orientation-based labeling with our adaptive clustering method. The results in Table 1 and Table 2 demonstrate the overall effectiveness of our approach. For the most direct comparison, Table 3 explicitly ablates the performance of our method against CosPlace on the same backbones (DINOv2 and ResNet50), clearly isolating and validating the significant contribution of our adaptive clustering.

Response to Q2:

This is a very interesting and insightful suggestion. We have not yet explored the application of our method in a continual learning setting, but we agree it is a promising direction. Using our adaptive clustering to refine labels over time as a system encounters new data could be a way to handle long-term appearance changes. We thank you for this idea and welcome further discussion on its potential.

Response to weakness1:

We thank you for highlighting this ambiguity. We would like to clarify: We do not propose a new architecture from scratch. Our contribution is a training framework, MutualVPR, which integrates adaptive clustering with a feature extractor in a mutual learning loop. As detailed in Section 3.1, this framework is a training methodology applied to classification-based architecture. Our novelty lies in the co-evolution of clustering and feature learning to resolve supervision inconsistency, not in the base architecture itself. We first demonstrate the effectiveness of our method by comparing it with current state-of-the-art approaches in Table 1, where our method consistently outperforms others across benchmarks.

To isolate the contribution of our adaptive clustering, we compare with CosPlace in Table 3, which uses a fixed labeling strategy. This serves as an ablation, showing the benefit of our dynamic labeling. Additionally, the case of K=1 in Table 4, equivalent to disabling our clustering, leads to clearly worse performance, further confirming the value of our method.

We acknowledge the importance of evaluating a broader set of baselines. In response, we have already included additional comparisons with both classification-based and contrastive learning methods. As reflected in Reviewer W9pq’s comments, our method still achieves state-of-the-art performance under the same feature dimensionality, demonstrating its robustness and generality. We will continue to incorporate more baseline comparisons in future revisions to further validate the effectiveness of our approach.

Response to weakness2:

We thank you for this point and agree that the fixed nature of K is a limitation, which we have acknowledged in the paper. We conducted an ablation study in Section 4.5.2 to analyze its impact.

Generally estimate the optimal K is currently infeasible and has to be decided by ablation. However, we found that the choice of an optimal K is related to the dataset's characteristics and the backbone.

On the SF-XL dataset, as regarding to the backbone, interestingly, the descriptors from ResNet50 exhibited more clearly defined cluster boundaries, while those from DINOv2 often showed overlapping clusters — in some cases, even two classes with different orientation labels appeared to belong to a single cluster. This suggests that ResNet50 tends to produce more distinct separations, whereas DINOv2 brings similar-looking images closer in the feature space.This observation corresponds well with our results in Table 4, where ResNet50 performs better with K=6, and DINOv2 performs better with K=3.

While these findings provide some insights, we do not yet have a definitive criterion for selecting the optimal K. Therefore, we still regard the choice of K as a limitation of our approach.

Response to weakness3:

We appreciate your comments and would like to clarify that on the SF-XL-Occlusion dataset, our method achieves the best performance, which strongly supports our core motivation of preserving label consistency. Detailed discussions can be found in Section 4.4 and Appendix B.1–B.2.

In contrast, EigenPlaces performs better than ours on the SF-XL-testv1 split. As noted in the main text (line 171 in the paper), we attribute this to overfitting. Upon further analysis, we find that SF-XL contains multi-view training samples. Following the EigenPlaces pipeline, images are cropped from panoramic views facing the same direction, which introduces strong viewpoint sensitivity. Since SF-XL-testv1 also consists primarily of multi-view test samples, this leads to overfitting.

However, EigenPlaces’ sampling strategy does not handle occlusion well. When the viewpoint is fixed and occluded by obstacles, label inconsistency arises — a core challenge our method addresses. This is evidenced by our 10-point performance gain over EigenPlaces on SF-XL-Occlusion, highlighting its inferior generalization ability.

Anyway, we agree that this reasoning should be made explicit, and we will include this analysis in the appendix for clarity.

Response to weakness4:

This is an insightful question. Taking contrastive learning — another representative paradigm in VPR — as an example, we can better understand the distinction between learning paradigms. In contrastive learning, samples are divided into positive and negative pairs, and the goal is to increase the distance between them. However, distances among positive samples are typically not explicitly constrained, which gives rise to a variety of hard sample mining strategies.

In contrast, classification-based methods enforce clear class boundaries, naturally encouraging samples within the same class to cluster more tightly, though the separation between different classes may be less structured. As demonstrated in Section 4.4 and Appendix B.1, our method leverages this structure to pull semantically similar samples (e.g., different views of the same place) closer in the feature space, thereby improving retrieval performance.

While our work is built upon the classification paradigm, we believe the core idea — dynamically aligning semantically similar instances — could be extended to enhance contrastive learning as well, for instance by informing hard negative mining with our learned clusters.

This is beyond the current scope of our study, but we consider it a promising direction for future research.

Response to sugesstions:

Thank you for your helpful suggestions. For suggestion1, we have addressed it in our response to Weakness 1, and we have added new baselines. The corresponding results are included in our response to Reviewer W9pq.

For suggestion2, we have responded under Weakness 3, and we will include the related discussion in the appendix.

We sincerely thank the reviewer for his insightful explanation regarding the strengths of EigenPlaces. We fully agree that the interpretation: linking EigenPlaces’ superior performance to its strong viewpoint-invariant inductive prior that aligns well with the SF-XL-testv1 benchmark’s structure, is accurate and much more precise than our earlier description. We appreciate the reviewer clarifying that this advantage reflects a form of task specialization rather than traditional overfitting, which helps deepen the understanding of the method’s behavior.

Our main argument centers around the different inductive priors imposed by class construction strategies during training:

EigenPlaces (geometry-driven) constructs classes by grouping views that observe the same focal point from different direction. This strategy imposes a strong viewpoint-invariant prior, which is particularly effective in benchmarks such as SF-XL-testv1, where the database entries differ significantly in viewing direction due to panorama cropping. The resulting performance reflects a favorable alignment between this inductive bias and the benchmark’s structural properties—what we initially (and incorrectly) referred to as “overfitting”.

Our Method (appearance-driven) forms classes via clustering in visual feature space, without relying on geometric information. While this does not explicitly enforce viewpoint invariance, the clusters are often semantically or visually coherent. Moreover, neighboring clusters typically correspond to spatially adjacent views, offering flexibility and resilience in complex scenarios involving occlusion, scene clutter, or subtle semantic variation.

To quantify this distinction, we sampled approximately 6k UTM grids (~0.7M images). For each grid, we performed k-means clustering based on both EigenPlaces and our method, and measured the feature distances between clusters of opposing headings (0° vs. 180°). On average, EigenPlaces exhibited the smallest distances (1.022), followed closely by ours (1.097), confirming its stronger viewpoint invariance, meaning that extreme viewpoints have closer feature distances and are thus more likely to be correctly recalled. This experiment confirms that EigenPlaces effectively encodes viewpoint invariance, explaining its superior performance on benchmarks with extreme viewpoint changes like SF-XL-testv1.

However, it is important to emphasize that our method is not devoid of viewpoint invariance; rather, it adopts a different inductive bias and is not specifically specialized for handling such extreme viewpoint variations.

Yet, as shown in Table 2 and Appendix B.2, our method demonstrates stronger performance in visually complex environments, such as those with heavy occlusion or clutter, thanks to its finer-grained and adaptive grouping of nearby viewpoints. This advantage is reflected in smaller feature distances under moderate viewpoint shifts, indicating better local viewpoint consistency and robustness to partial occlusion.

We believe this highlights a meaningful trade-off: EigenPlaces leverages a structured, geometry-based prior that enforces strong global invariance, whereas our method promotes adaptability through appearance-driven learning, which may be better suited for unstructured, occlusion-heavy, or cluttered scenes. Importantly, our method also effectively handles viewpoint variations, though with a different inductive bias than EigenPlaces.

Due to the submission policy, we are unable to provide visualization results at this stage. However, we promise to discuss this aspect and include the corresponding experiments in our final version.

We sincerely appreciate your clarification and the use of more precise terminology to help express our point more accurately than we initially did. We're also grateful that you found value in our work. If there are any remaining concerns, we'd be happy to further discuss them. We genuinely hope you enjoyed our submission and would kindly consider a higher score if you feel our clarifications have addressed your concerns.

Response to weakness1:

Thank you for this suggestion to broaden our experimental comparison. We have conducted additional experiments, and the results for GeM, Conv-AP, SALAD, and AnyLoc are presented below. Following Reviewer uVwD’s recommendation, we also include additional results on the Pitts250k-test dataset. Furthermore, in response to weakness2, we include results for CricaVPR (512-dimensional) under the same training and evaluation settings as our main method.

Our method continues to achieve state-of-the-art performance at an equivalent descriptor dimension and notably demonstrates a substantial advantage on the challenging SF-XL-Occlusion dataset.

For the baseline setups:

• GeM: We trained a ResNet50 backbone on the GSV-Cities dataset for 40 epochs.
• SALAD (512+32): We trained a DINOv2 backbone for 40 epochs on GSV-Cities to ensure consistency with our BoQ (512) training.
• CricaVPR: We applied PCA to reduce the descriptor dimensionality to 512 for a fair comparison.
• AnyLoc: The original implementation uses ViT-G/14 with VLAD or GeM. However, ViT-G/14 incurs slow inference on the SF-XL dataset. To maintain practical and comparable settings, we used ViT-B/14 with GeM for evaluation, which provides a balanced trade-off between performance and efficiency.

The results are as follows:

All methods will be included in the final manuscript, along with a detailed description of the experimental setup.

Since GCL and SFRS are difficult to scale for training on the GSV-Cities dataset, we do not include them in the main experiments to ensure fairness. However, in response to the reviewer’s suggestion, we present their testing results and experimental settings here for reference.

For GCL(ResNet50+GeM), we directly used the official pretrained weights and evaluated under our testing protocol. SFRS has a high feature dimensionality (32768), which makes it difficult to test on the SF-XL dataset. Therefore, we used the official pretrained weights and applied PCA to reduce the dimension to 4096. However, this dimensionality reduction led to a significant performance gap compared to the results reported in the original paper.

The results are as follows:

Response to weakness2:

Thank you for highlighting this important point.

Our primary baselines are CosPlace and EigenPlaces, both of which typically use 512-dimensional descriptors. To ensure consistency and fairness, we adopted the same dimensionality across our main comparisons.

For BoQ, which also uses DINOv2 as the backbone, we chose 512 dimensions to allow for a fair and consistent comparison under the same training and evaluation settings. Although CricaVPR also uses DINOv2, its performance is known to be sensitive to batch size. Since our evaluation uses a batch size of 1, we followed the official CricaVPR setup and retained the 4096-dimensional descriptors to match their published settings.

In response to your suggestion, we have now included CricaVPR (512) in our evaluation, and the results are provided in our response to weakness1.

For methods such as MixVPR, which use ResNet50, we retained their original feature dimensionality to preserve the integrity of their published configurations.

Additionally, we are currently training a version of our method using 4096-dimensional descriptors to explore performance at higher dimensions. However, due to time constraints during the rebuttal period, training is still in progress and results are not yet available. We will include these results and discussions in the final version of the manuscript.

Response to weakness3:

The omission of the GeM reference was our mistake. We have added GeM as a baseline and will include the proper citation.

Response to weakness4:

Thank you for this suggestion. Our use of LMCL classifier follows the precedent set by CosPlace, as we mentioned in the final paragraph of Section 3.2.2. Similarly, we employed GeM pooling as it is a standard and widely-adopted aggregation method in the VPR field.

We intentionally kept the descriptions of these modules concise, as our primary focus is on our mutual learning framework rather than established components. However, we agree with your that providing more detail would improve the paper's accessibility for readers who may not be deeply familiar with this specific domain.

Therefore, we will accept your valuable suggestion and add a more detailed description of both the LMCL classifier and GeM pooling to the appendix in the final manuscript.

Response to weakness5:

Thank you for catching the color inconsistency in Figure 2; we will correct it in the revision.

Response to weakness6:

We will add all relevant training parameters to the "Implementation Details". Thank you for your suggestion.

The author's detailed explanation and additional experiment results are appreciated and address most of my concerns. I have understood the inconvenience to train your model with 4096-dim descriptors due to time constraints of the rebuttal period, so I would expect related experimental results in the future revised version. I will maintain my original rating for this manuscript.

Response to Q1:

Thank you for raising this important point.

In Table 1, we followed the original implementations of CosPlace and EigenPlaces, both of which use ResNet50 as the backbone. To ensure a fair comparison, we additionally performed ablation studies in Table 3, where MutualVPR and CosPlace are compared using both ResNet50 and DINOv2 backbones under consistent data settings.

We acknowledge that our comparison does not include a version of EigenPlaces with a DINOv2 backbone. While we attempted such experiments, the results were notably suboptimal (e.g., Tokyo247: R@1 = 81.4, R@5 = 90.2; SF-XL-testv1: R@1 = 75.3, R@5 = 83.1). We suspect this may be due to the sampling strategy and fine-tuning procedures employed in EigenPlaces, which may not be compatible with the representational characteristics of DINOv2. Consequently, these results may not reflect the backbone’s true potential. To avoid possible misinterpretation, we opted not to include these results in the main paper. We will clarify this rationale in the final version for transparency.

Response to Q2:

Thank you for this valuable suggestion.

As fine-tuning methods are not the core focus of our work, we did not include a detailed comparison in the main paper. However, we had previously conducted experiments comparing MulConV with the PEFT method proposed in [1], and we ultimately chose MulConV due to its superior performance in our setting.

In response to your comment, we have now conducted additional experiments using the PEFT method from [2]. The following are our experimental results:

The results show that MulConV remains better suited to our framework, though other PEFT approaches also perform strongly. This highlights both the effectiveness and flexibility of our proposed framework in accommodating different fine-tuning strategies.

We will include the results and discussion of all these fine-tuning methods in the final version of the manuscript. Thank you again for your insightful feedback.

[1]Lu F, Zhang L, Lan X, et al. Towards seamless adaptation of pre-trained models for visual place recognition[J]. arXiv preprint arXiv:2402.14505, 2024.

[2]Jin T, Lu F, Hu S, et al. EDTformer: An Efficient Decoder Transformer for Visual Place Recognition[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2025.

Response to Q3:

For Q3, you are correct — our method does involve two forward passes. To clarify, the UTM-based geographic grid is predefined. At the start of each epoch, we sample a subset of UTM grid cells, each of which contains images captured from approximately the same location. For each selected grid cell, we group its images and perform clustering within the group to assign pseudo-labels. We extract features using the current model without gradients (i.e., the feature extractor is frozen during clustering), then assign labels based on clustering results. Once labels are updated, standard supervised training proceeds as usual.

Here is the corresponding pseudocode:

for epoch in epochs:
    sampled_grids = sample_grid_cells(dataset) 
    for grid in sampled_grids:
        images = get_images_from_grid(grid)
        with torch.no_grad():
            features = model(images)
        labels = cluster(features)
        update_labels(images, labels)
    for iteration in iterations:
        train_step()  # standard training with updated labels

The motivation is that at the early stage, clustering based on initial features may assign visually similar images to different classes or visually dissimilar images to same classes, leading to label inconsistency. By updating the pseudo labels progressively during training, we ensure that the grid-based labels and the learned feature space evolve together, resulting in mutual reinforcement.

We have conducted supporting experiments to validate our method. Specifically, we selected queries within the same geographic grid and visualized their assigned labels before and after training. Through multiple trials, we observed that some queries were assigned to different classes after training, indicating that label updates did occur. We discussed the benefits of this strategy in Section 4.4, and the corresponding visualizations can be found in Appendix B.1.

Response to weakness1:

Thank you for highlighting these important references! We will incorporate them into our Related Works section. Specifically, we will add [5] to our comparison results. Regarding [6], while it presents an interesting sample mining strategy, it is tailored for triplet construction in contrastive learning by mining positive samples. This approach is not directly applicable to our classification-based framework, which does not rely on explicit positive/negative sampling.

We will also include [1–4] in the revised manuscript, including [3] as the citation for the SF-XL Occlusion dataset and [4] for its relevance in parameter-efficient fine-tuning.

Response to weakness2:

We appreciate your constructive feedback on the manuscript’s clarity. We agree that certain parts—particularly the “Problem Analysis” section—can be improved for readability.

Since our clustering visualizations are based on class-level groupings, we believe it is appropriate to retain the term “Class.” However, to reduce confusion, we will clearly define “View X” and “Class Y” prior to our discussion on label inconsistency. Specifically: “View X” refers to the camera viewing direction derived from CosPlace’s heading-based orientation labels (i.e., fixed labels), indicated by Roman numerals. “Class Y” refers to the groups obtained through our clustering process.

We will revise this explanation in both the main text and Figure 2 to improve clarity for readers less familiar with the task. We also welcome further feedback on how to improve this section.

Additionally, we will adopt your suggestion to clarify that “manual labeling” refers specifically to using heading-based labels with fixed splits. If you deem it necessary, we are happy to relocate the Related Works section to follow the Introduction for better flow.

On Comparisons with SALAD and CliqueMining:

We thank the reviewer for emphasizing the importance of comparisons with SALAD and CliqueMining (CM). We have included SALAD results in our response to Reviewer W9pq’s Q1.

Regarding CM, we would like to clarify that it is a sample mining strategy specifically designed for contrastive learning. In the original CM paper, CM was used to mine training data from MSLS, and SALAD was subsequently trained on this mined dataset (MSLS + GSV-Cities). While we have not previously evaluated SALAD+CM ourselves, we agree this is a valuable comparison. However, it is worth noting that other contrastive learning baselines are evaluated on standard datasets without additional sample mining. Applying CM solely to SALAD could introduce inconsistencies in comparative evaluation.

We acknowledge SALAD+CM as a strong SOTA method. To provide a fair comparison, we retrained SALAD+CM with 512+32D embeddings using the original SALAD configuration. The results are as follows:

While SALAD+CM achieves higher R@1 on certain datasets, our method (MutualVPR) consistently achieves superior R@5 performance across most benchmarks. Notably, on the challenging SF-XL-Occlusion dataset, our approach exhibits a clear advantage in both R@1 and R@5, demonstrating robustness in visually complex and occluded scenarios.

Training CosPlace with DINOv2:

We trained CosPlace with the DINOv2 backbone and MulConv adapter—consistent with our method—to ensure a fair comparison and to isolate the effects of our training pipeline. Details are provided in Section 4.5.1.

On Differences from Reported CosPlace and EigenPlaces Results:

We re-trained CosPlace and EigenPlaces ourselves following their original configurations. Minor discrepancies from previously published results may stem from differences in image preprocessing (e.g., resolution, normalization). However, to ensure fair comparisons, all baselines in our experiments were evaluated under identical preprocessing pipelines. If further clarification is needed, we are happy to provide additional details.

Fine-tuning:

We appreciate the reviewer’s recognition of our fine-tuning approach. We agree that ablation studies are important and have discussed various strategies in Reviewer Xqf7’s Q2.

We experimented with fine-tuning the last DINOv2 layers using GeM pooling, testing both the last two and last four transformer blocks. However, we observed minimal improvement in training loss and significant degradation in validation performance—particularly when fine-tuning the last four layers, which resulted in faster and more severe performance drops. We attribute this to catastrophic forgetting: modifying deeper backbone layers tends to overwrite pre-trained representations, leading to instability and poor generalization.

In contrast, fine-tuning the last layers of ResNet50—used in both our method and CosPlace—yielded strong performance. We hypothesize this is due to ResNet50 being pretrained on supervised tasks, which produce more task-aligned and stable features. DINOv2, being self-supervised, learns more general-purpose representations that are more sensitive to parameter updates, making naive fine-tuning less effective. Similar instability was observed when fine-tuning the last few layers of DINOv2 in CosPlace and EigenPlaces.

Thus, we advocate for adapter-based fine-tuning in classification-based VPR methods, which better preserves stability and generalization performance.

If you have any further questions or concerns, we would be happy to discuss them with you.

Response to Q1:

We thank the reviewer for this insightful question, which allows us to clarify an important aspect of our classification-based approach to VPR. We believe there may be a slight misunderstanding regarding the role of the classifier and the use of labels during inference.

As noted in line 121 of the paper, the classifier is used only during training to guide the network in learning discriminative features. At inference time, we do not use the classifier or any form of "test labels". Instead, retrieval is performed directly using the learned feature descriptors via distance.

Regarding how our labeling strategy impacts recall@k, our dynamic geo-grid labeling ensures that images of the same place, even if captured from diverse and inconsistent viewpoints, may be assigned to different clusters during training. However, since these clusters are geometrically close, the learned descriptors for such images remain nearby in the embedding space. As a result, during retrieval, these images are still within the distance threshold and are correctly recalled.

This design enhances recall@k by increasing descriptor robustness to viewpoint variation. We have discussed this mechanism in line 203 of the paper and further supported it through visualizations and quantitative results in Appendix B.2.

Response to Q2:

Thank you for pointing this out. We clarify that the 131,072-dimensional NetVLAD descriptor reported in Table 1 is correct in our context. Specifically, we re-trained NetVLAD on the GSV-Cities dataset using a ResNet50 backbone (which outputs 2048-dimensional features) and the standard configuration of 64 cluster centers, resulting in a descriptor of size 2048 × 64 = 131,072. The original NetVLAD setup uses a VGG16 backbone with 512-dimensional features and 64 clusters, leading to a descriptor of 512 × 64 = 32,768. This explains the difference in reported dimensions.

Regarding evaluation on SF-XL-testv1, although NetVLAD descriptors can be precomputed offline, nearest-neighbor search requires loading over 2 million database descriptors into memory, which exceeded the hardware limits of our setup and prevented us from running the test.

As for Pitts250k-test, while we acknowledge its relevance, prior works such as SALAD [1] and CricaVPR [2] often report results on only one of Pitts30k or Pitts250k, given that performance trends are consistent across both. Following this convention, we chose to evaluate on Pitts30k-test. Nevertheless, we appreciate your suggestion and will include Pitts250k-test results in the final version of the paper. Preliminary results for these methods are provided in our response to Reviewer W9pq'Q1.

[1]Lzquierdo S, Civera J. Optimal transport aggregation for visual place recognition[C]/Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 17658-17668. [2]Lu F, Lan X, Zhang L, et al. Cricavpr: Cross-image correlation-aware representation learning for visual place recognition[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 16772-16782.

Response to Q3:

Thank you for this important question regarding our methodology for selecting the number of clusters, K. To clarify, we train our method only on the SF-XL dataset, using a single universal set of hyperparameters (including K, learning rate, etc.) across all experiments. Our ablation study on K (Table 4) deliberately varies both K and the backbone architecture to explore their impact on performance.

Interestingly, we observe that ResNet50-based descriptors tend to produce more clearly defined cluster boundaries, while DINOv2-based features often yield overlapping clusters—in some cases, even classes with different orientation labels appear to fall within a single cluster. This indicates that ResNet50 induces more separable features, whereas DINOv2 tends to pull visually similar images closer in feature space. This observation aligns with Table 4: ResNet50 performs better with K = 6, while DINOv2 performs better with K = 3.

The optimal value of K is also tied to the intrinsic structure of the training data. For example:

• SF-XL: This dataset is ideal for our classification task, containing many multi-view images per location. These exhibit viewpoint-induced supervision inconsistencies—precisely the challenge our method is designed to address. For this dataset, K = 6 or 3 is appropriate.
• GSV-Cities: This dataset lacks the above challenge, as all images from a location share the same orientation. Thus, K = 1 is sufficient and preferred.

In summary, the optimal choice of K is related to both the characteristics of the dataset and the feature space induced by the backbone. While these trends provide useful guidance, the exact K should still be validated through empirical ablation studies.

Response to weakness1:

Thank you for the suggestion. We are committed to strengthening our experimental comparisons. We will include comparisons with SALAD and AnyLoc in the final manuscript. Preliminary results for these methods are provided in our response to Reviewer W9pq.

Response to weakness2:

EigenPlaces performs better than our method on the SF-XL-testv1 split. As noted in the main text (line 171), we attribute this to overfitting. In the EigenPlaces training pipeline, images are cropped from panoramic views facing the same direction, introducing strong viewpoint diversity. Since SF-XL-testv1 also consists mainly of multi-view test samples, this leads to overfitting. However, EigenPlaces does not handle occlusion well. When a fixed viewpoint is occluded by obstacles, label inconsistency arises—a core challenge our method is designed to address. This is supported by our 10-point performance gain over EigenPlaces on SF-XL-Occlusion, demonstrating EigenPlaces’ inferior generalization capability under occlusion.

Response to weakness3:

Regarding the selection of the number of clusters K, we have provided a detailed discussion in our response to Q3. In future work, we plan to explore different datasets and investigate how to select an optimal or dynamic K, based on dataset characteristics and the structure of the feature space.