acceptodds
Under review as a conference paper at ICLR 2027

FM-ReID: Selective Competitive Token Routing for Object Re-Identification

14 views

Abstract

Object re-identification (ReID) faces a recurring challenge: different identities can share highly similar global appearances, while the cues that distinguish them are localized, heterogeneous, and visible only under particular viewpoints. This challenge arises in animal ReID through markings, contours, and scars, in person ReID through subtle clothing and accessory cues, and in vehicle ReID through localized appearance details. Although visual foundation models encode such information in dense tokens, a single holistic descriptor can obscure discriminative local signals. We propose FM-ReID, an end-to-end framework that formulates local representation learning as selective competitive token routing. Its Competitive Fine-grained Mining module uses multiple mining queries and a residual query to compete for dense DINOv3 tokens. Above-prior selection retains tokens preferentially allocated to each mining query, while the residual slot receives tokens excluded from the retrieval descriptors. The resulting multi-query descriptors are jointly trained with a holistic representation for retrieval, without fixed spatial partitions or equal-area constraints. FM-ReID achieves strong results on animal, person, and vehicle ReID benchmarks, supporting competitive token routing as an effective way to augment holistic foundation-model representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Object re-identification, Fine-grained representation learning, Competitive token routing, Metric learning

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

This paper proposes FM-ReID for object ReID, which uses a DINOv3 backbone and lets K mining queries plus one residual query compete for patch tokens by softmax across queries, keeping only above-prior assignments. The final descriptor concatenates the holistic feature with K fine-grained features, trained with ReID losses and a diversity loss. It reports good results on animal, person and vehicle benchmarks (79.8 mAP on MSMT17, 85.03 mTop-1 on WildlifeReID-10k), and provides a self-built fish benchmark FM-FISH.

Critical strengths
  1. The competition design is reasonable: making queries fight for each token instead of each looking at all tokens alone is a simple and effective way to stop several local features from learning the same region. 2、Results are good on many benchmarks — animal, person, vehicle — with small extra cost, and the ablation is careful: one variable changed at a time, same backbone and same training recipe.
Critical weaknesses

1、The residual query is one of the three headline designs, but their own ablation shows it adds only 0.49 and 0.1 points. 2、Above-prior selection is the key step of the method, but it is never tested alone. Without this ablation, and without comparing to normal per-query softmax, we cannot know if the "competition" story is what brings the gain. 3、Much of the gain may come from the stronger DINOv3 backbone, not the method: a plain DINOv3 baseline already beats TransReID on both tables. Table 2 has no same-backbone row, so backbone gain and CFM gain are mixed. 4、The comparison is not fair in feature size: their descriptor is 4608-d while all baselines use 768-d. Part of the gain may just come from concatenating more dimensions, and Table 5 does not report gallery or distance cost.

Questions for authors

1、Table 4 shows that the residual query only increases by 0.49/0.1 points. Could you report the situation using multiple seeds and variance, and compare it with the version using only prior selection? 2、The prior selection was not eliminated (Equation 4). Did using the ordinary competitive softmax alone already reach the reported values? 3、Could a row for DINOv3+GAP be added in Table 2, and report how much gain is left compared to TransReID under the same 768-d constraint conditions? 4、The descriptor is 4608-d, compared to the 768-d baseline. What would happen if the fused features were projected back to the 768-d space? 5、When a mining query almost does not obtain any tokens, formula 4 normalizes the weights approaching zero through ε. Will this amplify the noise on occluded images? How frequently does the empty support situation occur during training? 6、Will FM-FISH be publicly released, including fixed branches and an official evaluation server?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Keog (Q1–Q3)

Thank you for your careful review and constructive suggestions. We appreciate your recognition of the competitive allocation design, the results across animal, person, and vehicle ReID benchmarks, and our controlled ablation study. Your questions have helped us strengthen the experimental analysis and clarify the contribution of each component.

Q1.

With competition and above-prior selection retained, adding the residual query improves WildlifeReID-10k mTop-1 from 84.54% to 85.03% and BAKS from 82.90% to 83.33%, giving gains of 0.49 and 0.43 percentage points, respectively.

Following your suggestion, we conducted additional experiments on MSMT17 to examine the contribution of the residual query. We compared Full with a version that removes the residual query while retaining competition and above-prior selection. The table below summarizes the paired results for random seeds 1234, 271, and 307 as mean ± sample standard deviation:

SettingmAP (%)Rank-1 (%)
Full79.45 ± 0.1991.26 ± 0.23
Without residual query79.38 ± 0.1991.15 ± 0.29

Across these comparisons, average mAP and Rank-1 gains are 0.0704 and 0.1086 percentage points. The sample variances are 0.0374 and 0.0508 for Full, compared with 0.0376 and 0.0816 without the residual query, respectively, in squared percentage points.

Our motivation is to provide an additional competitive destination for background or other evidence that the mining descriptors need not include. The residual query receives no explicit background supervision, and its output is excluded from the retrieval descriptor.

We believe that a dedicated background query could separate background evidence from identity-discriminative information more explicitly. The current residual query can serve as an initial template for this idea, providing an auxiliary competitive route that could support more targeted background modeling and help mining queries focus on informative identity cues.

Q2.

To isolate the contribution of above-prior selection, we removed it while retaining competitive allocation, the residual query, and the same supervision strategy.

SettingmAP (%)Rank-1 (%)Rank-5 (%)Rank-10 (%)
Competitive softmax without selection79.668591.465895.737296.7321
Independent per-query attention without selection79.686291.405895.728696.6978
Full79.675191.517395.814496.8608

The results show that competitive softmax without selection already achieves comparable accuracy, while above-prior selection provides marginal numerical improvements in Rank-1, Rank-5, and Rank-10.

Importantly, the two mechanisms serve different purposes: competitive allocation distributes visual evidence across queries, whereas above-prior selection explicitly filters weak assignments before feature aggregation. The variant without selection still retains the residual query and coordinated identity supervision. Thus, these results should not be interpreted as evaluating softmax alone independently of the proposed framework. Our contribution lies in the integration of competitive allocation, selective aggregation, and coordinated supervision, rather than the accuracy gain of any single component.

Q3.

We report CLS as the DINOv3 backbone baseline and CLS+GAP as an additional global control. CLS+GAP learns a fusion of CLS and average-pooled patch features. Both output 768-dimensional descriptors and contain no CFM. Entries are mAP / Rank-1 (%):

RepresentationMSMT17Market-1501DukeMTMC-reIDOcc-Duke
CLS75.97 / 89.6691.49 / 96.1183.94 / 91.9763.21 / 70.77
CLS+GAP77.48 / 90.5791.39 / 95.7284.27 / 91.2963.99 / 71.13

For the same-dimensionality comparison on MSMT17, PCA was fitted separately using each method's training features only:

MethodDimensionmAP (%)Rank-1 (%)
TransReID + DINOv3 + PCA76872.5387.64
FM-ReID + DINOv3 + PCA76879.4091.26

TransReID retains JPM and camera SIE, with the same DINOv3 initialization and shared training recipe. At 768 dimensions, FM-ReID improves mAP and Rank-1 over this TransReID control by 6.87 and 3.62 percentage points, respectively, and over CLS by 3.43 and 1.60 points. We will add the DINOv3 CLS+GAP control to Table 2.

AuthorsOpenReview ↗

Response to Reviewer Keog (Q4–Q6)

We continue our response with Q4–Q6.

Q4.

Retrieval dimensions vary across established multi-branch ReID configurations:

Method and configurationCompositionDimension
PCB/RPP, G descriptor6 × 204812,288
PCB/RPP, H descriptor6 × 2561,536
MGN8 × 2562,048
TransReID, ViT-B + JPM5 × 7683,840
CLIP-ReID, ViT-B/16768 + 5121,280
CLIP-ReID, ResNet-502048 + 10243,072
FM-ReID6 × 7684,608

(https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/Yifan_Sun_Beyond_Part_Models_ECCV_2018_paper.pdf), MGN, TransReID, and CLIP-ReID, Table 11.

To examine compression directly, we fitted PCA using training features only, without using query or gallery features to fit the projection:

MSMT17DimensionmAP (%)Rank-1 (%)
Full460879.6891.52
Full + PCA76879.4091.26

The compressed descriptor retains most of the retrieval performance at 768 dimensions. Together with the matched-dimensionality comparison in Q3, these results support the value of the learned representation beyond its original descriptor size.

Q5.

The normalization keeps the pre-projection token aggregate bounded when a mining query has little or no support. Mining features also enter retrieval with reduced weights, preserving a dominant holistic contribution.

The denominator includes both the total retained weight and a small stabilizing constant, epsilon. When support is empty, all retained weights are zero, so the corresponding head's token aggregate is also zero. For nonempty support, the normalized weights sum to at most one, preventing an unbounded aggregation magnitude caused solely by a small retained-weight sum. Other heads, output projection, and the query's residual and refinement paths can still produce a nonzero complete descriptor.

Each retrieval branch is L2-normalized. Before concatenation, the holistic feature has weight 1 and each miner has weight 0.25. With five nonzero normalized miners, the holistic term accounts for approximately 76.2% of the normalized similarity weighting, limiting the contribution of potentially unreliable mining features.

We also sampled every 20 forward passes throughout 120 training epochs:

DatasetSampled unitsEmpty-support rate
Occ-Duke1,589,7609.47%
Market-15011,390,0809.93%
FM-FISH3,073,28010.75%

A sampled unit is an image × attention head × mining query, excluding the residual query. These percentages concern individual heads, not failed images or complete descriptors. No sampled unit had a positive retained-weight sum at or below epsilon, which is set to 1e-8. All sampled aggregation weights and attention outputs were finite.

These observations support numerical stability in the recorded training runs. Numerical boundedness alone does not guarantee that every retained feature is informative under occlusion; the reduced miner weights limit its contribution to the final similarity.

Q6.

Upon acceptance, we will publicly release FM-FISH through an accessible online download link, together with the FM-ReID code and reproduction materials, including dataset construction details, fixed data splits, identity mappings, training configurations, evaluation protocols, and standardized evaluation scripts.

Reproducible evaluation will be supported through the released scripts for local execution. Given the deployment and maintenance requirements of an online evaluation server, we do not currently provide this service, but may consider introducing it as the project develops.

We sincerely thank the reviewer for the constructive questions and suggestions, which have helped us strengthen our experimental analysis and clarify our contributions.

We would like to emphasize that our main contribution lies in a unified feature-mining and supervision framework. FM-ReID combines competitive token allocation, selective feature aggregation, and coordinated identity supervision to learn complementary identity representations from dense foundation-model features. The additional experiments, including controlled ablations and matched-dimensionality comparisons, further support the effectiveness of the overall framework.

We hope these clarifications and additional results adequately address your concerns, and we greatly appreciate your time and consideration.

AuthorsOpenReview ↗

Response to Reviewer Keog

Dear Reviewer,

Thank you again for your constructive feedback. As the discussion period is approaching its end, we would greatly appreciate your feedback on our rebuttal and additional experiments. Please let us know if any concerns remain that we could further clarify.

Thank you for your time and consideration.

I thank the authors for the response, which partially addresses my questions. After also considering the other two reviews, I decide to keep my original score.

Reviewer_2ywdOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

FM-ReID introduces a Competitive Feature Mining (CFM) module built on DINOv3 to extract complementary local identity features for object re-identification. Learnable queries compete to select informative visual tokens, while diversity regularization encourages distinctive local representations. Experiments across animal, person, and vehicle ReID benchmarks demonstrate its effectiveness, although the method increases feature dimensionality and does not reduce visual-token computation.

Critical strengths

Promising benchmark results, excellent writing clarity.

Critical weaknesses
  1. Insufficient Novelty: The novelty of this work is severely limited, as it largely combines existing techniques. Local feature partitioning and dynamic query-based feature extraction have been extensively explored since the CNN era of ReID, as exemplified by PCB and OSNet. Learnable tokens and query-aware mechanisms have also been widely studied across various tasks, particularly in prompt tuning. Furthermore, orthogonality-based feature regularization is a well-established strategy. The paper does not sufficiently demonstrate a fundamentally novel contribution beyond these existing approaches.
  2. Insufficient Presentation and Experimental Analysis: The overall framework is overly simplistic, and the paper lacks sufficient figures, experiments, and visualization analyses to support its claims. The experimental analysis is also too limited to provide deeper insights into the proposed method. In addition, the paper suffers from disorganized formatting and inadequate presentation of information, making the methodology and contributions difficult to assess.
  3. Efficiency and Scalability: Given the increased feature dimensionality, parameter count, and memory usage, does FM-ReID offer practical advantages for large-scale ReID applications?
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2ywd

Thank you for recognizing the experimental results and writing clarity. We hope the following clarifications and additional experiments address your concerns.

Q1.

Our contribution concerns how dense foundation-model features are allocated and selected to form complementary identity descriptors.

PCB uses uniform spatial partitions, while RPP refines part membership through learned soft assignments. CFM couples token assignments through competition across queries, introduces a residual allocation route excluded from the retrieval descriptor, and selectively retains assignments through an above-prior rule. This allows unequal query supports and permits tokens to remain outside the mining descriptors. OSNet instead combines convolutional features at multiple scales through channel-wise gates.

We acknowledge the competitive allocation principle of Slot Attention. CFM develops this principle for identity retrieval by combining selective mining with holistic identity context. Identity-classification and triplet losses supervise the holistic feature, each concatenated holistic–mined pair, and their joint fusion. The contribution lies in this feature-mining and supervision framework, beyond the use of learnable queries or regularization alone.

PCB and OSNet are included in Table 2, with published MSMT17 mAP / Rank-1 of 40.40 / 68.20 and 52.90 / 78.70, respectively, compared with 79.68 / 91.52 for FM-ReID. These methods use different backbones; the matched-DINOv3 comparisons below examine our framework's contribution more directly.

Q2.

Figures 2–3 present the framework and competitive token routing. Figure 5 visualizes the holistic and individual miner responses.

To distinguish the contribution of the framework from the backbone, we compared global features and Full using the same DINOv3 backbone and matched training settings:

MSMT17mAP (%)Rank-1 (%)
DINOv3 CLS75.9789.66
DINOv3 CLS+GAP77.4890.57
FM-ReID Full79.6891.52

Full improves mAP over CLS and CLS+GAP by approximately 3.7 and 2.2 percentage points, respectively. These comparisons support the value of learning supplementary local identity features on top of the same backbone.

We also adapted TransReID to the same DINOv3 initialization and shared training recipe. It achieves 72.69% mAP and 87.95% Rank-1, compared with 79.68% and 91.52% for FM-ReID in that comparison. Additional attention controls and hyperparameter sweeps examine the mining mechanism and parameter sensitivity.

We would appreciate specific examples of the formatting issues and any remaining gaps in the analysis so that we can address them directly.

Q3.

We evaluated both descriptor compression and resource costs. Multi-branch descriptors are also used by TransReID (ViT-B+JPM: 3840 dimensions) and CLIP-ReID, Table 11 (ViT-B/16: 1280; ResNet-50: 3072).

For FM-ReID, we fitted PCA using training features only:

MSMT17DimensionmAP (%)Rank-1 (%)
DINOv3 CLS76875.9789.66
FM-ReID Full460879.6891.52
FM-ReID Full + PCA76879.4091.26

At 768 dimensions, FM-ReID retains gains of 3.43 points in mAP and 1.60 in Rank-1 over CLS.

For 82,161 gallery images with FP32 descriptors:

DescriptorGallery storageMedian distance-computation time
Full, 4608 dimensions1.514 GB503.6 ms
PCA, 768 dimensions0.252 GB82.1 ms

Timing uses four CPU threads and 64 queries against the full gallery, excluding encoding, projection, and ranking. PCA reduces gallery storage to one-sixth while retaining most retrieval accuracy.

Table 5 reports encoding costs separately. At 256 × 128 resolution, FP32, and batch size 1, Full adds approximately 1.41% FLOPs, 9.66% inference parameters, and 9.28% peak GPU memory over CLS+GAP. CFM aggregates backbone outputs; compression reduces retrieval storage and computation while retaining the original encoder.

Dear Reviewer,

Please respond to the authors’ rebuttal on OpenReview. Briefly indicate whether the rebuttal addresses your concerns and confirm or revise your score accordingly.

Best regards, Area Chair

Reviewer_2ywdOpenReview ↗

Response to Authors

Thank you for the additional results. Most of my concerns have been addressed. However, the novelty of the work remains limited. Nevertheless, in recognition of your efforts, I am willing to increase my score to 2.

AuthorsOpenReview ↗

Response to Reviewer 2ywd

Thank you for your reconsideration and for acknowledging our additional experiments and clarifications. We appreciate your constructive feedback and your willingness to revise the score. Your comments will help us further improve the presentation and clarify the contributions of our work.

Reviewer_dbt8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Based on the DINOv3 backbone model, this paper proposes FM-ReID. Its core proposition is the Competitive Fine-grained Mining module (CFM), which includes K learnable queries and a residual slot, applies Softmax normalization along the query dimension, and uses a hard threshold to retain only tokens whose assigned weights exceed a uniform prior, which capable of suppressing background and non-discriminative patches before aggregation, while the residual slot absorbs the discarded information. The paper also contributes FM-FISH, a fine-grained fish ReID benchmark dataset (45,632 images, 724 identities), and reports results on animal, person, and vehicle ReID benchmarks.

Critical strengths
  1. The CFM module builds on the classical Slot Attention competitive allocation mechanism, combining a residual slot with hard-threshold truncation above the uniform prior. This design allows different queries to focus on local regions at different scales and avoids forcibly including all background tokens into the foreground feature descriptor. I believe this indeed addresses a genuine failure mode of slot attention.
Critical weaknesses
  1. The performance improvement relative to the DINOv3 baseline is modest, and Table 4 does not separately isolate the residual competitive mechanism, so I cannot determine whether this mechanism is indeed the key factor for improvements in complex scenarios. Supplementary material is also missing, and the description of the new dataset is overly brief, making evaluation difficult.

  2. In Figure 4(b) main text (line 427) and Table 2, the MSMT17 results for FM-ReID are 79.8% mAP and 91.3% Rank-1; however, the last row of Table 4, which appears to be the same complete model configuration, reports 79.7% mAP and 91.4% Rank-1. Given the small overall margin of improvement, this discrepancy is significant.

Questions for authors
  1. Are these two MSMT17 results from the same run? Please clarify the differences between the two settings and specify which set of values should be considered the final results.

  2. In the comparison between Table 1 and Table 3, if the baseline methods (e.g., TransReID) were retrained with the DINOv3 backbone under the same training scheme, would FM-ReID still maintain a clear advantage?

  3. Could the paper include hyperparameter settings and corresponding tuning experiments to allow evaluation of the method’s sensitivity?

1 comment
AuthorsOpenReview ↗

Response to Reviewer dbt8

Thank you for your careful review and for recognizing the value of CFM's competitive allocation and selective aggregation in addressing background interference.

Q1.

The main-table result used weight decay 0.007, whereas the Full result in the ablation table used 0.005. We overlooked this difference when preparing the tables.

We have now unified weight decay at 0.005 and rerun Full and the component ablations under the same training configuration. The updated Full result is 79.68% mAP and 91.52% Rank-1 on MSMT17. We will update the main table, ablation table, and corresponding text in the revised manuscript consistently.

Q2.

Following your suggestion, we replaced TransReID's backbone with the same pretrained DINOv3 used by FM-ReID, retaining its JPM branches and camera SIE. Both methods used the same shared training settings and MSMT17 evaluation protocol.

MethodDimensionmAP (%)Rank-1 (%)
DINOv3 CLS76875.9789.66
DINOv3 CLS+GAP76877.4890.57
TransReID + DINOv3384072.6987.95
FM-ReID + DINOv3460879.6891.52

FM-ReID improves over this TransReID comparison by 6.99 percentage points in mAP and 3.57 in Rank-1. Its mAP gains over CLS and CLS+GAP are approximately 3.7 and 2.2 points. CLS+GAP learns a fusion of CLS and average-pooled patch features; neither global baseline contains CFM.

We also repeated the global-baseline comparisons with three additional seeds. Full improves mean mAP over CLS and CLS+GAP by 3.17 and 2.03 points, with positive paired differences in all three runs.

For a comparison at the same retrieval dimension, we fitted PCA separately using each method's training features only:

MethodDimensionmAP (%)Rank-1 (%)
TransReID + DINOv3 + PCA76872.5387.64
FM-ReID + DINOv3 + PCA76879.4091.26

FM-ReID retains its advantage at 768 dimensions, supporting the benefit of learning local retrieval features alongside the global DINOv3 representation. Table 4 reports the component ablation.

Q3.

The existing query-number sweep covers K=1K=1 to 1616 under the original weight decay of 0.007. Increasing KK from 1 to 5 raises mAP from 78.2% to 79.8%. For K=4K=4 to 1616, mAP ranges from 79.4% to 80.0%, and Rank-1 from 91.3% to 91.8%. Five queries provide accuracy close to the best observed result with a relatively compact descriptor.

We also varied the diversity-loss coefficient under the updated training configuration, keeping all other settings fixed:

λdiv\lambda_{\mathrm{div}}mAP (%)Rank-1 (%)
079.7591.47
0.02579.7891.77
0.0579.6891.52
0.179.7091.34

The mAP range is approximately 0.11 percentage points and the Rank-1 range is 0.43 points, suggesting limited sensitivity to this coefficient within the tested range.

Our main contribution is a feature-mining framework that learns complementary identity information from dense visual representations. CFM couples token assignments through competition across queries, provides an additional allocation route through the residual query, and selectively aggregates evidence into local descriptors that complement the holistic representation.

The supervision strategy is also integral to this design. We apply identity-classification and triplet losses to the holistic feature, each concatenated holistic–mined feature pair, and the joint fusion of all retrieval branches. Each mining branch receives supervision with holistic identity context, while the fusion objective supervises the combined representation. Together, selective mining and coordinated supervision connect local evidence extraction with identity discrimination at multiple levels. We hope these comparisons and analyses help clarify the contribution of the complete framework.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal and the revision

We thank the reviewers and the AC for their time and constructive feedback. We appreciate the recognition of competitive allocation (dbt8, Keog), the results across ReID benchmarks (Keog, 2ywd), and writing clarity (2ywd). We summarize our contribution and additional evidence below.

Our main contribution is a general feature-mining and supervision framework for object ReID. FM-ReID learns complementary identity information from dense visual representations, with a common design evaluated across person, animal, and vehicle benchmarks.

CFM learns query-dependent local representations without predefined anatomical parts. Competition across queries couples their assignments to each token, while an auxiliary residual query provides an additional allocation route whose output is excluded from the retrieval descriptor. This flexible aggregation accommodates identity cues that vary in location, shape, and scale across object categories.

Our supervision strategy trains these local features together with the holistic representation. Identity-classification and triplet objectives supervise the holistic feature, each concatenated holistic–mined pair, and their joint fusion. Each miner therefore learns with holistic identity context, while fusion supervision coordinates the combined retrieval representation. This connects local evidence extraction with identity discrimination at both branch and joint-representation levels.

In the reported comparisons, FM-ReID achieves higher mAP than the listed prior methods on all four person ReID benchmarks, including the occluded setting. Together with our animal ReID results, these findings support the value of complementary local identity features and their coordinated supervision.

The following table summarizes the additional evidence addressing the reviewers' main concerns. Detailed settings and results are provided in the corresponding responses.

ConcernAdditional evidence or clarificationResponse
Gains attributable to DINOv3Under the same DINOv3 initialization and shared training recipe, adapted TransReID with JPM and camera SIE achieves 72.69% mAP / 87.95% Rank-1, compared with 79.68% / 91.52% for FM-ReID.dbt8 Q2; Keog Q3; 2ywd Q2
Efficiency and descriptor dimensionalityDescriptor concatenation is also used by TransReID and CLIP-ReID. Encoding adds approximately 1.41% FLOPs over CLS+GAP. PCA fitted only on training features compresses 4608 to 768 dimensions, with small losses of approximately 0.27 percentage points in mAP and 0.26 in Rank-1.Keog Q3–Q4; 2ywd Q3
CFM mechanismsComponent comparisons examine removal of the residual query, competitive softmax without selection, and independent per-query attention to assess the allocation and aggregation choices.Keog Q1–Q2
Empty-support behaviorAn aggregation bound and sampled training statistics address numerical stability; all recorded aggregation weights and attention outputs were finite.Keog Q5
ReproducibilityUpon acceptance, we will release FM-FISH, fixed splits, code, training configurations, and evaluation scripts for local evaluation.Keog Q6

Our contribution combines flexible local feature mining, coordinated identity supervision, and evaluation across animal, person, and vehicle ReID. The global controls, matched-backbone comparisons, and compact-descriptor results support the framework's strong retrieval performance beyond the backbone alone. We hope these clarifications help the AC and reviewers assess its methodological contribution and practical value. Thank you for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.