- 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
- 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
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 configuration | Composition | Dimension |
|---|---|---|
| PCB/RPP, G descriptor | 6 × 2048 | 12,288 |
| PCB/RPP, H descriptor | 6 × 256 | 1,536 |
| MGN | 8 × 256 | 2,048 |
| TransReID, ViT-B + JPM | 5 × 768 | 3,840 |
| CLIP-ReID, ViT-B/16 | 768 + 512 | 1,280 |
| CLIP-ReID, ResNet-50 | 2048 + 1024 | 3,072 |
| FM-ReID | 6 × 768 | 4,608 |
(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:
| MSMT17 | Dimension | mAP (%) | Rank-1 (%) |
|---|---|---|---|
| Full | 4608 | 79.68 | 91.52 |
| Full + PCA | 768 | 79.40 | 91.26 |
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:
| Dataset | Sampled units | Empty-support rate |
|---|---|---|
| Occ-Duke | 1,589,760 | 9.47% |
| Market-1501 | 1,390,080 | 9.93% |
| FM-FISH | 3,073,280 | 10.75% |
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.
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.
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:
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.
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 (%):
For the same-dimensionality comparison on MSMT17, PCA was fitted separately using each method's training features only:
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.