acceptodds
Under review as a conference paper at ICLR 2027

USIL: Uncertainty-Aware Selective Invariant Learning for Low-Light Visible-Infrared Person Re-Identification

Abstract

Low-light visible–infrared person re-identification (VI-ReID) aims to retrieve the same pedestrian across heterogeneous cameras under severe illumination degradation. The coexistence of reliable identity structures and uncertain cross-spectral observations can contaminate invariant representations when no cue-wise distinction is made. To address this problem, we propose a novel Uncertainty-Aware Selective Invariant Learning (USIL) framework. USIL consists of an Uncertainty-Aware Selective Invariant Modeling mechanism that selectively propagates stable identity cues and estimates location-wise uncertainty, and an Uncertainty-Guided Low-Light Enhancement strategy that uses the resulting reliability to regulate feature compensation and descriptor aggregation. Furthermore, Distributional Cross-modal Retrieval-Risk Regularization is introduced to constrain the upper tail of batch-wise ranking violations through Bayesian-quadrature-inspired Dirichlet reweighting. Extensive experiments on three standard VI-ReID benchmarks demonstrate consistent improvements over the reproduced IRL baseline, including gains of 0.77% Rank-1 and 1.10% mAP on SYSU-MM01 All-Search.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visible-infrared person re-identification, low-light representation learning, identity-cue uncertainty, selective state-space modeling, distributional retrieval-risk regularization

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_MDR7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Uncertainty-Aware Selective Invariant Learning (USIL) for visible-infrared person re-identification under low-light conditions. USIL consists of three main components: Uncertainty-Aware Selective Invariant Modeling (USIM), Uncertainty-Guided Low-Light Enhancement (UGLE), and Distributional Cross-modal Retrieval-Risk Regularization (DCRR). USIM estimates spatial reliability using bidirectional selective state-space modeling, while UGLE uses reliability estimates to modulate feature compensation and feature aggregation. DCRR introduces a distributional cross-modal retrieval-risk regularizer to emphasize potentially problematic ranking cases during training. The proposed method is evaluated on SYSU-MM01, RegDB, and LLCM. The reported results improve over the IRL baseline across the evaluated benchmarks and retrieval settings.

Critical strengths
  1. This paper addresses the VI-ReID challenge that local features vary in identity relevance under illumination degradation and cross-spectral mismatch. Reusing a shared reliability representation for feature compensation and descriptor aggregation provides a well-motivated and coherent design.

  2. By combining spatial reliability estimation (USIM/UGLE) with sample-level ranking risk (DCRR), the proposed method attempts to link local feature quality to downstream retrieval errors rather than optimizing enhancement in isolation.

Critical weaknesses
  1. Key benchmark results do not validate the complete method. Appendix E.2 states that the reported RegDB and LLCM results are pre-fix runs that “do not validate post-fix DCRR results.” Since the corrected procedure adds local reliability supervision, these results cannot establish that the complete method consistently outperforms IRL across all three benchmarks.

  2. The tail-risk motivation of DCRR is insufficiently justified. The proposed method applies symmetric Dirichlet weights to a sorted set of sample risks before computing an upper quantile of the resulting weighted aggregates. Because symmetric Dirichlet weights are exchangeable, sorting the risks does not change the distribution of the weighted sum. Although the paper establishes basic properties of this surrogate, it does not sufficiently explain why this construction specifically controls the high-risk tail of individual sample risks or demonstrate its advantages over established alternatives such as CVaR-based objectives.

  3. Missing required AI use statement: The paper omits the required “AI use statement” section specified in the submission format. The authors should include this statement to ensure compliance with the submission requirements.

Questions for authors
  1. Have the corrected DCRR method, including local reliability supervision, been evaluated on RegDB and LLCM? If so, what are the results compared with IRL under the same protocol?

  2. What empirical evidence supports the claim that the proposed objective effectively focuses on high-risk samples? How does it compare with standard alternatives such as top-k or CVaR-based risk minimization?

  3. Could the authors report the end-to-end parameter count, FLOPs, training time, and inference latency of USIL compared with IRL, ideally including a parameter-matched baseline?

  4. Beyond the existing module-level ablations and hyperparameter sensitivity analyses, can the authors isolate the effects of the L+ term, local reliability supervision, and the risk-feedback pathway on retrieval performance?

5 comments
AuthorsOpenReview ↗

Response to Reviewer MDR7

We sincerely thank the reviewer for recognizing the motivation behind USIL, particularly its shared spatial reliability modeling and the connection between local identity-cue quality and cross-modal retrieval risk. We appreciate the constructive concerns regarding cross-dataset validation, the mathematical interpretation of DCRR, computational efficiency, component attribution, and AI-use disclosure. We address each point below.

Q1. Corrected Cross-Dataset Validation (W1 + Q1)

We thank the reviewer for highlighting the inconsistency between the pre-fix and corrected configurations in Appendix E.2. After rechecking the training records, we distinguish two configurations: USIL, which disables explicit stripe-level reliability supervision, and USIL⁺, which enables it. Both retain the same USIM–UGLE–DCRR architecture and otherwise matched training settings. Importantly, USIL already performs spatial reliability modeling and reliability-guided enhancement; USIL⁺ introduces an additional explicit supervision constraint on the relative reliability of body regions rather than a new spatial modeling module.

Additional Cross-Dataset Evaluation. To directly assess the corrected training procedure, we report USIL⁺ results on SYSU-MM01, RegDB, and LLCM while retaining the original USIL results for transparency.

Table R1. Comparison of USIL and USIL⁺ across six retrieval settings (Rank-1 / mAP, %).

Dataset / ProtocolUSILUSIL⁺Improvement (pp)
SYSU-MM01 All-search77.00 / 73.1880.04 / 75.18+3.04 / +2.00
SYSU-MM01 Indoor-search83.12 / 85.5885.73 / 86.73+2.61 / +1.15
RegDB VIS→IR90.98 / 82.4491.70 / 83.45+0.72 / +1.01
RegDB IR→VIS88.54 / 80.0891.55 / 81.74+3.01 / +1.66
LLCM VIS→IR64.38 / 66.4866.36 / 68.46+1.98 / +1.98
LLCM IR→VIS55.63 / 61.9156.75 / 63.25+1.12 / +1.34

Evidence and Scope. USIL⁺ consistently improves Rank-1 and mAP over USIL across all six retrieval settings and also outperforms the reproduced IRL baseline on these metrics. These controlled comparisons provide additional empirical support for the effectiveness of explicit stripe-level reliability supervision across all three benchmarks. We note that some SYSU-MM01 Rank-10 values decrease relative to USIL; therefore, our conclusion is specifically restricted to the consistent Rank-1 and mAP improvements, without claiming universal improvements across every metric or independent training seed.

Manuscript Revisions. We retain both configurations in Tables 1–2, clarify their distinction in the table notes, and update their training configurations in Appendix E.2. The original component-wise ablation in Table 3 remains unchanged to avoid conflating configuration-level improvements with the existing module-level attribution.

AuthorsOpenReview ↗

Response to Reviewer MDR7

Q2. Mathematical Validity and Empirical Comparison of DCRR (W2+Q2)

We thank the reviewer for this rigorous observation. We agree that symmetric Dirichlet weights are exchangeable; therefore, sorting the batch risks neither changes the distribution of their weighted aggregate nor preferentially assigns larger expected weights to higher-risk samples.

Conservative Aggregate-Risk Formulation. Importantly, Eq. (11) includes an additional worst-case pseudo-risk ((rmax⁡=1r_{\max}=1)). For (NN) observed risks, the implemented aggregation is

+w_{N+1}r_{\max}, \qquad \mathbf w\sim\operatorname{Dirichlet}(\mathbf1_{N+1}),$$ with the exact conditional expectation $$\mathbb E[Q] =\frac{\sum_{i=1}^{N}r_i+1}{N+1} =\bar r+\frac{1-\bar r}{N+1}.$$ This introduces a nonnegative upward shift relative to the empirical mean. Nevertheless, ($L^{+}$) regularizes an upper quantile of stochastic aggregate retrieval risks, rather than the upper tail or CVaR of the original sample-risk distribution. Accordingly, we do not claim a per-sample tail-risk guarantee. **Retrieval-Oriented Optimization.** Eq. (9) constructs differentiable cross-modal ranking-violation risks, while the non-detached ($L^{+}$) branch propagates gradients through pairwise similarities to identity descriptors. Together with the separate risk-prediction and reliability-feedback objectives, this provides a retrieval-oriented regularization pathway without implying guaranteed preferential optimization of the hardest samples. **Controlled Empirical Comparison.** To evaluate the practical contribution of Dirichlet aggregation, we compare five risk-aggregation configurations on RegDB (VIS→IR, epoch 140): no aggregation, mean-risk, top-($k$), $\mathrm{CVaR}_{0.9}$, and Dirichlet aggregation. The no-aggregation configuration disables only ($L^{+}$) while retaining the remaining risk-related objectives, corresponding to the Full USIL w/o ($L^{+}$) configuration in our independent DCRR ablation (Q4). As summarized in Table R2, Dirichlet achieves 90.98% Rank-1 and 82.44% mAP, compared with 90.73%/82.28% without aggregation, yielding additional gains of +0.25/+0.16 percentage points. It also outperforms $\mathrm{CVaR}_{0.9}$ (90.70%/81.64%), top-($k$) (90.55%/81.68%), and mean-risk (90.41%/81.51%). The independently evaluated empirical hard-risk CVaR is 0.2005 for Dirichlet, lower than the reported values of 0.2365–0.2374 for the three alternative aggregation objectives. These results support the empirical effectiveness of Dirichlet aggregation under the evaluated protocol without establishing theoretical control over individual risk tails. Table R2. Controlled comparison of risk aggregation objectives on RegDB (VIS→IR, epoch 140). | Risk aggregation | Rank-1 (%) ↑ | mAP (%) ↑ | Hard-risk CVaR ↓ | | ----------------------------------- | ------------ | --------- | ---------------- | | No aggregation (($\lambda^+=0$)) | 90.73 | 82.28 | 0.2376 | | Mean-risk | 90.41 | 81.51 | 0.2365 | | Top-$k$ risk | 90.55 | 81.68 | 0.2373 | | $\mathrm{CVaR}_{0.9}$ | 90.70 | 81.64 | 0.2374 | | $\color{red}{\textbf{Dirichlet (Ours)}}$ | **$\color{red}{90.98}$** | **$\color{red}{82.44}$** | **$\color{red}{0.2005}$** | Note: Retrieval metrics are averaged over the ten official RegDB train/test splits. Empirical hard-risk CVaR is reported to four decimal places. The no-aggregation configuration disables ($L^{+}$) while retaining the remaining risk-related objectives. **Manuscript Revisions.** We clarify Dirichlet exchangeability, the exact expectation, and the scope of stochastic aggregate-risk regularization in Sec. 3.4 and Appendix B.5. We also revise the Abstract, Introduction, contribution statement, and Conclusion to avoid unsupported per-sample tail-risk guarantees. The empirical risk-aggregation comparison is summarized in Sec. 4.3, with the complete results provided in Table R2 of this rebuttal. The incremental effect of ($L^{+}$) is independently summarized in Q4.
AuthorsOpenReview ↗

Response to Reviewer MDR7

Q3. Architectural Effectiveness and Full-Model Computational Efficiency (Q3)

We thank the reviewer for requesting a comprehensive evaluation of computational efficiency. USIM employs bidirectional selective state propagation to jointly model spatial identity cues and estimate their reliability for subsequent feature enhancement and aggregation. Theorem 1 establishes (O(HW)O(HW)) scaling with spatial sequence length under fixed feature and state dimensions. However, this operator-level property does not imply that the complete USIL model is faster than IRL.

Controlled Architectural Comparison. To further assess the effectiveness of this design, we replace only the USIM context operator with residual CNN, spatial attention, and lightweight Transformer alternatives while retaining the remaining components and matched training/evaluation protocols. As shown in Table R4, bidirectional Mamba achieves the highest Rank-1 and mAP on both SYSU-MM01 All-search and RegDB VIS→IR.

Table R4. Controlled comparison of spatial modeling operators within USIL (Rank-1 / mAP, %).

Spatial Modeling OperatorSYSU-MM01 All-searchRegDB VIS→IR
Residual CNN73.42 / 70.5889.22 / 79.61
Spatial Attention75.41 / 71.2490.59 / 81.66
Lightweight Transformer76.10 / 72.2290.29 / 82.14
Bidirectional Mamba (Ours)\color{red}{\textbf{Bidirectional Mamba (Ours)}}77.00/73.18\color{red}{77.00 / 73.18 }90.98/82.44\color{red}{ 90.98 / 82.44}

Note: All variants retain the same backbone, reliability head, UGLE, DCRR, and training/evaluation protocols, replacing only the USIM context operator.

Compared with the strongest non-Mamba alternative for each metric, bidirectional Mamba improves Rank-1 and mAP by 0.90 and 0.96 percentage points on SYSU-MM01 All-search, respectively. On RegDB VIS→IR, the corresponding improvements are 0.39 and 0.30 percentage points, relative to spatial attention for Rank-1 and the lightweight Transformer for mAP. These results support the effectiveness of bidirectional Mamba relative to the evaluated alternatives without implying universal architectural superiority.

Full-Model Computational Overhead. Beyond the architectural comparison, we evaluate the reproduced IRL baseline and full USIL under matched inference settings. As shown in Table R5, USIL increases the parameter count from 181.223M to 181.349M (+0.126M, +0.07%) and FLOPs from 41.584G to 42.446G (+0.862G, +2.07%). Inference latency increases from 25.79 to 28.46 ms (+2.67 ms, +10.35%), while throughput decreases from 38.77 to 35.14 FPS (−3.63 FPS, −9.36%). These measurements quantify the computational overhead of the complete USIL inference model relative to IRL.

Table R5. Computational overhead of full USIL relative to the reproduced IRL baseline.

MetricIRLFull USILΔ (USIL − IRL)
Params (M) ↓181.223181.349+0.126 (+0.07%)
FLOPs (G) ↓41.58442.446+0.862 (+2.07%)
Inference (ms) ↓25.7928.46+2.67 (+10.35%)
FPS ↑38.7735.14−3.63 (−9.36%)

Note: Computational metrics are reported for the complete inference models under matched evaluation settings. Absolute differences are calculated as Full USIL minus IRL, and relative changes are computed with respect to IRL.

Accuracy–Efficiency Trade-off and Parameter Matching. Full USIL introduces only a marginal parameter increase of 0.126M (+0.07%) and an additional 0.862G FLOPs (+2.07%) compared with IRL. Its inference latency increases by 10.35%, while throughput decreases by 9.36%. Thus, the improved retrieval performance comes with modest additional parameter and arithmetic costs and a measurable runtime trade-off, rather than an unconditional inference-speed advantage. The closely matched parameter counts provide a parameter-scale-comparable comparison with IRL, although the two models retain different computational operations.

Training-Time Scope. We do not infer training-time efficiency from parameter counts or FLOPs without matched wall-clock measurements.

Manuscript Revisions. We clarify the architectural motivation in Sec. 3.2 and summarize the controlled operator comparisons and measured full-model accuracy–efficiency trade-off in Sec. 4.3. This distinguishes the effectiveness of bidirectional Mamba within USIL from the additional computational cost of the complete framework.

Reviewer_3QkBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Uncertainty-Aware Selective Invariant Learning (USIL) for low-light visible-infrared person re-identification. The framework consists of three components: Uncertainty-Aware Selective Invariant Modeling (USIM), Uncertainty-Guided Low-Light Enhancement (UGLE), and Distributional Cross-modal Retrieval-Risk Regularization (DCRR). These components aim to estimate local identity-cue reliability, guide feature enhancement and aggregation, and reduce cross-modal retrieval risks. Experiments on three public benchmarks demonstrate improvements over the reproduced IRL baseline.

Critical strengths
  1. The paper is well organized and clearly presented. The overall framework and individual modules are described with detailed mathematical formulations and architectural illustrations, making the proposed approach relatively easy to understand.

  2. The paper provides useful qualitative analyses and implementation details. The retrieval visualizations, reliability maps, and feature-distance distributions offer additional insights into the model's behavior. The supplementary material also provides detailed algorithmic descriptions and implementation settings, which are helpful for understanding and reproducing the method.

Critical weaknesses
  1. The rationale for the research and the necessity of the task are not sufficiently explained. The paper’s analysis of the unique challenges of low-light VI-ReID compared to traditional VI-ReID, as well as the limitations of existing methods, remains insufficiently in-depth and does not fully demonstrate the necessity of designing a specialized method for this task.

  2. The experimental analysis in Figure 1 lacks sufficient persuasiveness. Figure 1 only compares IRL and USIL; we recommend including other representative VI-ReID methods to validate the performance of different existing methods under low-light degradation conditions, thereby strengthening the persuasiveness of the research motivation.

  3. There is a lack of sufficient comparison with recent state-of-the-art methods. Some of the methods used for comparison in the current experiments are relatively outdated, and representative visible-infrared pedestrian re-identification methods from the past two years are underrepresented. It is recommended to include more recent relevant methods related to low-light VI-ReID, Mamba-based ReID, or uncertainty modeling.

  4. The performance contribution of the DCRR module is relatively limited. According to the ablation experiments in Table 3, adding the DCRR module improves Rank-1 by only 0.16 percentage points and mAP by only 0.02 percentage points.

  5. There is a lack of thorough analysis regarding the necessity of using the Mamba architecture. The paper employs a bidirectional Mamba for local identity feature modeling but has not sufficiently explained its unique advantages over CNNs, Transformers, or traditional attention mechanisms.

6.There is a lack of analysis regarding the model’s computational complexity and runtime efficiency. The paper introduces bidirectional Mamba and several additional learning modules but fails to adequately assess the computational overhead resulting from these designs. It is recommended to supplement the analysis with metrics such as the number of model parameters (Params), computational load (FLOPs), and inference speed (FPS).

Questions for authors

See the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 3QkB

Q1. Motivation and Low-Light Evidence (W1+W2)

Motivation and Task Necessity. We thank the reviewer for these constructive suggestions. Low-light VI-ReID presents more than an enlarged cross-modal gap: illumination degradation selectively suppresses identity-relevant color and texture while potentially preserving misleading background, reflection, and sensor responses, resulting in spatially non-uniform identity-cue reliability. Unlike approaches primarily addressing modality discrepancy or feature enhancement, USIL explicitly couples local reliability estimation with feature enhancement, descriptor aggregation, and retrieval-risk feedback.

Experimental Evidence and Figure 1. To substantiate the motivation, we report the complete six-level exposure evaluation on SYSU-MM01 (Table R1), with all results averaged over ten gallery trials. Both methods exhibit performance degradation as illumination decreases, while USIL consistently outperforms the reproduced IRL baseline in Rank-1 and mAP at every evaluated exposure level. At the strongest darkening level (0.20), Rank-1/mAP improve from 62.77%/55.48% to 64.75%/57.58%, corresponding to gains of 1.98/2.10 percentage points. Accordingly, we clarify in the Introduction that Figure 1 provides a controlled comparison of illumination sensitivity against the reproduced IRL baseline, rather than a comprehensive robustness ranking across existing methods.

Table R1. Six-level visible-light exposure evaluation on SYSU-MM01 (averaged over ten gallery trials).

Exposure factorIRL Rank-1 / mAP (%)USIL Rank-1 / mAP (%)Gain Rank-1 / mAP (pp)
0.2062.77 / 55.4864.75 / 57.58+1.98 / +2.10
0.3670.31 / 64.4671.42 / 66.01+1.11 / +1.55
0.5273.23 / 68.1573.66 / 69.04+0.43 / +0.89
0.6874.76 / 70.1975.12 / 70.96+0.36 / +0.77
0.8475.89 / 71.5576.29 / 72.49+0.40 / +0.94
1.0076.23 / 72.0877.00 / 73.18+0.77 / +1.10

All metrics are averaged over ten gallery trials under the corresponding exposure setting. Gains are differences between averaged results, not statistical significance estimates across independent training runs.

Manuscript Revisions and Scope. We have clarified the distinctive challenges of low-light VI-ReID and the scope of Figure 1 in the revised Introduction. Together with the existing USIM–UGLE ablations in Table 3 and the reliability analyses in Appendix E.4, the exposure experiments provide complementary evidence for reliability-aware retrieval under illumination degradation. We restrict our conclusions to the evaluated settings without claiming universal robustness or probabilistic calibration.

AuthorsOpenReview ↗

Response to Reviewer 3QkB

Q2. Recent State-of-the-Art Comparisons (W3)

We thank the reviewer for this suggestion. We have expanded Related Work to discuss recent advances, including the WHU-MARS benchmark (CVPR 2026), recent Mamba-based person ReID, and relevant low-light VI-ReID studies. We also retain comparisons with methods for which SYSU-MM01, RegDB, or LLCM results are reported under the corresponding benchmark protocols, including the reproduced IRL baseline (ACM MM 2025). We distinguish the methodological relevance of recent studies from numerical comparisons across potentially different datasets and evaluation settings.

Methodological Distinction and Manuscript Revisions. WHU-MARS emphasizes unified retrieval across modalities and viewpoints, whereas USIL focuses on spatially non-uniform identity-cue reliability under low illumination. More specifically, USIL couples local reliability modeling with low-light enhancement and cross-modal retrieval-risk feedback. We clarify these distinctions in the revised Related Work while retaining the existing experimental results.Our empirical conclusions are therefore restricted to comparisons conducted under the reported benchmark protocols.

Manuscript Revisions and Scope. We have expanded the discussion of recent VI-ReID and Mamba-based approaches in Sec. 2 and clarified the corresponding benchmark comparisons in Sec. 4.2. Tables 1–2 now report both USIL and USIL$^{+}$, including additional results on RegDB and LLCM, while Appendix E.2 documents their training configurations. We restrict comparisons to the reported benchmark protocols and do not claim superiority over methods evaluated under incompatible settings.

[1] Zhao Y, Zhou Z, Yang B, et al. WHU-MARS: A Multispectral Aerial-Ground Benchmark Towards Any-Scenario Person Re-Identification[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 25461-25471.

[2] Wang D, Xing G, Liu Y. Low-light invariant representation learning for visible-infrared person re-identification[C]//Proceedings of the 33rd ACM International Conference on Multimedia. 2025: 8645-8653.

AuthorsOpenReview ↗

Response to Reviewer 3QkB

Q3. DCRR Contribution and Ablation (W4)

Incremental Contribution and Attribution. We thank the reviewer for highlighting the modest improvement of DCRR in Table 3. To further examine its contribution, we conduct additional ablations on SYSU-MM01 (All-search and Indoor-search) and RegDB (VIS→IR). We distinguish the combined contribution of DCRR-associated objectives from the isolated effect of the Dirichlet aggregate-risk regularizer L+L^{+}. As shown in Table R2, full USIL achieves the highest Rank-1 and mAP across all three retrieval settings.

Table R2. Ablation of DCRR-associated objectives across three retrieval settings.

Dataset / ProtocolConfigurationRank-1 (%) ↑mAP (%) ↑
SYSU-MM01 All-searchA. USIM + UGLE76.8473.16
B. Full USIL w/o L+L^{+}76.9273.17
C. Full USIL77.0073.18
SYSU-MM01 Indoor-searchA. USIM + UGLE82.1184.80
B. Full USIL w/o L+L^{+}82.7085.07
C. Full USIL83.1285.58
RegDB VIS→IRA. USIM + UGLE90.3982.16
B. Full USIL w/o L+L^{+}90.7382.28
C. Full USIL90.9882.44

Configuration A disables all DCRR-associated objectives; B retains the remaining risk-related objectives while disabling L+L^{+}; C uses the complete risk-related objective.

Independent Contribution of L+L^{+}. As summarized in Table R3, the complete risk-related configuration (A→C) improves Rank-1/mAP by +0.16/+0.02, +1.01/+0.78, and +0.59/+0.28 percentage points on SYSU All-search, SYSU Indoor-search, and RegDB, respectively. More importantly, isolating L+L^{+} (B→C) yields additional gains of +0.08/+0.01, +0.42/+0.51, and +0.25/+0.16 points, respectively. These consistently positive increments provide empirical evidence that L+L^{+} contributes beyond the remaining risk-related objectives, particularly under the SYSU Indoor-search protocol. We do not interpret these gains as proof of cross-seed statistical significance or certified per-sample tail-risk control.

Table R3. Performance gains of the complete DCRR-related objectives and the isolated L+L^{+} regularizer (percentage points).

Dataset / ProtocolA→C: Complete DCRR (R1 / mAP)B→C: Isolated L+L^{+} (R1 / mAP)
SYSU All-search+0.16 / +0.02+0.08 / +0.01
SYSU Indoor-search+1.01 / +0.78+0.42 / +0.51
RegDB VIS→IR+0.59 / +0.28+0.25 / +0.16

All differences are reported in percentage points. A→C measures the combined contribution of DCRR-associated objectives, while B→C isolates the incremental contribution of L+L^{+} with the remaining objectives retained.

Manuscript Revisions. We clarify the combined scope of the original DCRR ablation in Sec. 4.3 and the corresponding loss decomposition in Appendix C. We also summarize the additional positive contribution of L+L^{+} in Sec. 4.3.

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

This paper studies Visible-Infrared Person Re-Identification (VI-ReID) under low-light conditions and introduces an uncertainty-aware framework called USIL. The model uses a bidirectional selective state-space module (USIM) to estimate spatial reliability of identity cues, applies this estimate to guide low-light feature compensation and aggregation (UGLE), and adds a Dirichlet-based retrieval-risk regularizer (DCRR) during training. While connecting local cue reliability to global retrieval risk is an intuitive idea, the gains over the reproduced IRL baseline on three benchmarks are quite marginal. More importantly, I have major reservations regarding the mathematical validity of the DCRR objective, the circular validation of the uncertainty estimates, and inconsistencies in the experimental records.

Critical strengths
  1. Connecting spatially varying cue reliability under poor illumination to global retrieval errors offers a sensible perspective on robust cross-modal learning. Besides, using the same estimated reliability scores to guide both state-space feature compensation and final descriptor aggregation is a clean, unified design.
  2. The method is evaluated across three VI-ReID benchmarks under bidirectional retrieval settings, supplemented by illumination-degradation tests and qualitative visualizations that provide a solid starting point for evaluating the framework.
Critical weaknesses
  1. In Eq. (11), the risk surrogate is built by taking a symmetric Dirichlet-weighted average of sorted sample risks. Because symmetric Dirichlet weights are exchangeable, sorting the sample risks beforehand has no effect on the distribution of the weighted sum. Moreover, the expectation of this sum is simply the sample mean, and its variance shrinks as the batch size grows. Taking an upper quantile over these random convex combinations does not actually optimize the upper tail of the original sample-risk distribution. To justify DCRR, the authors need to clarify the math and compare it directly against standard alternatives like mean-risk, CVaR, or top-k hard-example mining.
  2. While the authors show that the predicted reliability correlates well with local identity consistency, the reliability head is explicitly trained using a closely related stripe-level ranking loss. Thus, this diagnostic does not independently prove that the module captures robust identity cues under illumination degradation. Besides, the ablation studies do not clearly disentangle the gains of selective state-space modeling and reliability-guided modulation from much simpler baselines, such as standard spatial attention or confidence-weighted pooling. Furthermore, the improvement over the reproduced IRL baseline is quite modest. In Table 3, adding DCRR only edges up Rank-1 by 0.16% and mAP by 0.02%, a margin that could easily fall within random seed variance.
Questions for authors
  1. Since symmetric Dirichlet weights are permutation-invariant, why does sorting the risks in Eq. (11) matter, and how does DCRR actually target high-risk samples? Also, how does it compare empirically against standard mean-risk, CVaR, and top-k risk?
  2. Can you validate the reliability maps outside the training objective, e.g., by testing whether local scores drop under synthetic occlusion, noise, or darkening? How does the model perform if you replace this module with simple spatial attention or constant/shuffled maps?
  3. Which code version (pre-fix or corrected DCRR) produced Tables 1 and 2? Please provide updated RegDB and LLCM results using the finalized code, along with multi-seed mean ±\pm std on SYSU-MM01 to verify that DCRR's +0.16% Rank-1 gain is statistically significant.
4 comments
AuthorsOpenReview ↗

Response to Reviewer CZRh

Q1. Mathematical Validity and Empirical Comparison of DCRR (W1+Q1)

We thank the reviewer for the rigorous observation. We agree that symmetric Dirichlet weights are exchangeable; therefore, sorting the batch risks neither changes the distribution of their weighted aggregate nor preferentially assigns larger expected weights to high-risk samples.

Conservative Aggregate-Risk Formulation. Importantly, Eq. (11) includes an additional worst-case pseudo-risk (rmax⁡=1r_{\max}=1). For NN observed risks, the implemented random aggregate is

Q=∑i=1Nwir(i)+wN+1rmax⁡,w∼Dirichlet⁡(1N+1),Q=\sum_{i=1}^{N}w_i r_{(i)}+w_{N+1}r_{\max},\qquad\mathbf w\sim\operatorname{Dirichlet}(\mathbf1_{N+1}),

with the exact conditional expectation

E[Q]=∑i=1Nri+1N+1=rˉ+1−rˉN+1.\mathbb E[Q]=\frac{\sum_{i=1}^{N}r_i+1}{N+1}=\bar r+\frac{1-\bar r}{N+1}.

Thus, the construction introduces a nonnegative upward shift relative to the empirical mean. Nevertheless, L+L^+ penalizes an upper quantile of stochastic Dirichlet-weighted aggregate risks, rather than the upper tail or CVaR of the original per-sample risk distribution. Accordingly, it does not establish a per-sample tail-risk guarantee.

Retrieval-Oriented Optimization. Eq. (9) constructs differentiable cross-modal ranking-violation risks, while the non-detached L+L^+ branch propagates gradients through pairwise similarities to identity descriptors. This provides a retrieval-oriented regularization pathway without guaranteeing preferential optimization of the hardest individual samples.

Controlled Empirical Validation. We compare five risk-aggregation configurations on RegDB (VIS→IR, epoch 140): no aggregation, mean-risk, top-kk, CVaR0.9\mathrm{CVaR}_{0.9}, and Dirichlet aggregation. As shown in Table R1, Dirichlet achieves 90.98% Rank-1 and 82.44% mAP, compared with 90.73%/82.28% without aggregation, corresponding to additional improvements of 0.25 and 0.16 percentage points, respectively.

Dirichlet also outperforms CVaR0.9\mathrm{CVaR}_{0.9} (90.70%/81.64%), top-kk (90.55%/81.68%), and mean-risk (90.41%/81.51%). Moreover, it achieves the lowest reported empirical hard-risk CVaR (0.2005), compared with 0.2365–0.2376 for the alternatives. These results provide empirical support for Dirichlet aggregation under the evaluated protocol without implying theoretical control over individual risk tails or cross-seed statistical significance.

Table R1. Controlled comparison of risk-aggregation objectives on RegDB (VIS→IR, epoch 140).

Risk aggregationRank-1 (%) ↑mAP (%) ↑Hard-risk CVaR ↓
No aggregation (λ+=0\lambda^+=0)90.7382.280.2376
Mean-risk90.4181.510.2365
Top-kk risk90.5581.680.2373
CVaR0.9\mathrm{CVaR}_{0.9}90.7081.640.2374
Dirichlet (Ours)\color{red}{\textbf{Dirichlet (Ours)}}90.98\color{red}{90.98}82.44\color{red}{82.44}0.2005\color{red}{0.2005}

Note: Retrieval metrics are averaged over the ten official RegDB train/test splits. Empirical hard-risk CVaR is reported to four decimal places. The no-aggregation configuration disables L+L^+ while retaining the remaining risk-related objectives. Its Rank-1/mAP results are also reported as the Full USIL w/o L+L^+ configuration in the separate DCRR ablations presented in our responses to Reviewer 3QkB (Q3, Table R2) and Reviewer MDR7 (Q4, Table R6). Manuscript Revisions. We clarify Dirichlet exchangeability, the exact expectation, and the scope of the stochastic aggregate-risk objective in Sec. 3.4 and Appendix B.5. We also revise the Abstract, Introduction, contribution statement, and Conclusion to avoid unsupported tail-risk guarantees. The risk-aggregation comparison is summarized in Sec. 4.3, with the complete results provided in Table R1 of this rebuttal. The incremental contribution of L+L^+ is further examined in the separate DCRR ablations.

AuthorsOpenReview ↗

Response to Reviewer CZRh

Q2. Reliability Validation and Simpler Alternatives (W2+Q2)

We thank the reviewer for emphasizing the importance of reliability validation beyond the training objective. On 64 visible–infrared identity pairs from held-out SYSU-MM01 identities, a frozen epoch-140 checkpoint achieves a Spearman correlation of 0.703 between decoded reliability and local identity margins, with a top–bottom reliability-group margin difference of 0.228. These findings indicate an association between learned reliability and local identity discrimination on unseen identities. To complement the held-out identity-margin analysis, we further evaluate reliability responses under controlled local corruptions using intervention-based criteria independent of the stripe-margin target.

Independent Local Corruption-Response Validation. To directly address the reviewer's concern, we additionally evaluate frozen-model reliability maps under controlled local occlusion, Gaussian noise, and darkening, each at three severity levels (Table R2). We measure the reliability decrease inside the corrupted region (ΔsA\Delta s _A), outside it (Δsout\Delta s _{\mathrm{out}}), their difference (Δssel\Delta s _{\mathrm{sel}}), and corruption-localization AUROC. Across all nine evaluated settings, ΔsA\Delta s _A and Δssel\Delta s _{\mathrm{sel}} are positive, and AUROC ranges from 0.60 to 0.80. For example, at occlusion severity 0.30, ΔsA=0.0190\Delta s _A=0.0190, Δsout=0.0030\Delta s _{\mathrm{out}}=0.0030, and AUROC reaches 0.80. These measurements provide complementary, intervention-based evidence that reliability decreases more strongly in locally corrupted regions. They are independent of the stripe-margin evaluation target, although the analysis remains limited to the tested synthetic corruptions and does not establish calibrated reliability probabilities.

Table R2. Local corruption-response validation using frozen USIL reliability maps.

CorruptionSeverityΔsA\Delta s_A ↑Δsout\Delta s_{\mathrm{out}}Δssel\Delta s_{\mathrm{sel}} ↑AUROC ↑
Occlusion0.10+0.0060+0.0015+0.00450.65
Occlusion0.20+0.0120+0.0020+0.01000.73
Occlusion0.30+0.0190+0.0030+0.01600.80
Noise0.10+0.0040+0.0010+0.00300.62
Noise0.20+0.0090+0.0015+0.00750.70
Noise0.30+0.0150+0.0020+0.01300.78
Darkening0.80+0.0030+0.0008+0.00220.60
Darkening0.50+0.0080+0.0012+0.00680.69
Darkening0.20+0.0140+0.0020+0.01200.77

Note: Positive ΔsA\Delta s_A denotes a decrease in reliability after corruption. We define Δssel=ΔsA−Δsout\Delta s_{\mathrm{sel}}=\Delta s_A-\Delta s_{\mathrm{out}}. AUROC evaluates the spatial discrimination of corrupted versus uncorrupted regions using reliability decreases. The reported responses are specific to the evaluated perturbations and are not probabilistic calibration metrics.

Table R3. Six-level visible-light darkening evaluation on SYSU-MM01 (ten-gallery-trial averages).

Exposure factorIRL Rank-1 / mAP (%)USIL Rank-1 / mAP (%)Gain (pp)
0.2062.77 / 55.4864.75 / 57.58+1.98 / +2.10
0.3670.31 / 64.4671.42 / 66.01+1.11 / +1.55
0.5273.23 / 68.1573.66 / 69.04+0.43 / +0.89
0.6874.76 / 70.1975.12 / 70.96+0.36 / +0.77
0.8475.89 / 71.5576.29 / 72.49+0.40 / +0.94
1.0076.23 / 72.0877.00 / 73.18+0.77 / +1.10

Note: All values are ten-gallery-trial averages. Gains are differences between averaged results, not evidence of statistical significance across independent training runs.

AuthorsOpenReview ↗

Response to Reviewer CZRh

Robustness Under Illumination Degradation. We also evaluate retrieval under six visible-light exposure levels on SYSU-MM01, averaged over ten gallery trials (Table R3). USIL outperforms the reproduced IRL baseline in both Rank-1 and mAP at every evaluated level. Under the strongest darkening (exposure factor 0.20), Rank-1/mAP improve from 62.77%/55.48% to 64.75%/57.58%, corresponding to gains of 1.98/2.10 percentage points. These results demonstrate improved retrieval performance under global darkening and complement, rather than replace, the local reliability-response validation.

Constant and Shuffled Reliability Maps. To evaluate the importance of the learned spatial structure, we compare learned, constant, and shuffled reliability maps using the same frozen checkpoint and evaluation split (Table R4). Learned maps achieve 90.39% Rank-1 and 82.38% mAP, compared with 86.07%/77.99% using constant maps. Three shuffled-map seeds yield Rank-1 scores of 85.49%–86.41% and mAP scores of 76.78%–77.62%. These inference-time interventions support the importance of the learned reliability distribution and its spatial correspondence, while not determining the performance of independently retrained replacement models.

Table R4. Frozen-checkpoint comparison of reliability-map replacements.

Reliability MapRank-1 (%) ↑mAP (%) ↑
Learned (Ours) \color{red}{\textbf{Learned (Ours) }}90.39\color{red}{90.39}82.38\color{red}{82.38}
Constant86.0777.99
Shuffled (seed 2027)85.6376.78
Shuffled (seed 2028)86.4177.62
Shuffled (seed 2029)85.4977.41

Note: These comparisons use the same frozen checkpoint and evaluation split, with no retraining after map replacement.

Comparison with Simpler Attention Mechanisms. We additionally compare the USIM-based configuration against Spatial Attention and Lightweight Attention alternatives (Table R5). The USIM-based configuration achieves 90.98% Rank-1 and 82.44% mAP, compared with 89.66%/81.38% for Spatial Attention and 89.90%/81.69% for Lightweight Attention. Relative to the strongest alternative for each metric, the corresponding improvements are 1.08 and 0.75 percentage points in Rank-1 and mAP, respectively. These results indicate an empirical advantage for the USIM-based configuration under the reported evaluation settings, without establishing cross-seed statistical significance or universal architectural superiority.


Table R5. Comparison with simpler reliability generators.

Reliability GeneratorRank-1 (%) ↑mAP (%) ↑
USIM (Ours) \color{red}{\textbf{USIM (Ours) }}90.98\color{red}{90.98}82.44\color{red}{82.44}
Spatial Attention89.6681.38
Lightweight Attention89.9081.69

Note: Results are reported for the evaluated reliability-generation alternatives. This comparison concerns the reported generator configurations and should be distinguished from the separate spatial-context-operator replacement experiment discussed in the responses to Reviewers 3QkB and MDR7.

Manuscript Revisions and Scope. Together with the existing USIM–UGLE component ablations in Table 3, these experiments provide complementary evidence at the local reliability, retrieval robustness, and architectural levels. We clarify these findings in the revised Introduction, Sec. 4.3, and Appendix E.4. Importantly, the controlled corruption experiment measures spatial reliability responses without reusing the stripe-margin diagnostic as its evaluation target. We nevertheless restrict our conclusions to the evaluated synthetic interventions and retrieval protocols, without claiming probabilistic calibration or universal robustness.

AuthorsOpenReview ↗

Response to Reviewer CZRh

Q3. Code-Version Consistency and Statistical Reliability (W2+Q3)

We thank the reviewer for emphasizing the importance of consistent experimental records. After rechecking the training configurations, we distinguish USIL from USIL⁺. Both configurations retain the same USIM–UGLE–DCRR architecture and otherwise matched training settings within each dataset, while USIL⁺ additionally enables explicit stripe-level reliability supervision. Specifically, USIL uses λrelloc=0\lambda_{\mathrm{rel}}^{\mathrm{loc}}=0, whereas USIL⁺ uses λrelloc>0\lambda_{\mathrm{rel}}^{\mathrm{loc}}>0. We retain the original results and report the additional supervised configuration separately in Tables 1–2 and Appendix E.2.

Additional Cross-Dataset Validation. To address the reviewer's concern regarding the finalized training procedure, we evaluate USIL⁺ on SYSU-MM01, RegDB, and LLCM under their respective benchmark protocols. The results are summarized in Table R6.

Table R6. Comparison of USIL and USIL⁺ across six retrieval settings (Rank-1 / mAP, %).

Dataset / ProtocolUSILUSIL⁺Improvement (pp)
SYSU-MM01 All-search77.00 / 73.1880.04 / 75.18+3.04 / +2.00
SYSU-MM01 Indoor-search83.12 / 85.5885.73 / 86.73+2.61 / +1.15
RegDB VIS→IR90.98 / 82.4491.70 / 83.45+0.72 / +1.01
RegDB IR→VIS88.54 / 80.0891.55 / 81.74+3.01 / +1.66
LLCM VIS→IR64.38 / 66.4866.36 / 68.46+1.98 / +1.98
LLCM IR→VIS55.63 / 61.9156.75 / 63.25+1.12 / +1.34

Note: The two configurations retain the same architecture and other training settings within each dataset; USIL⁺ additionally enables explicit stripe-level reliability supervision. SYSU-MM01 and LLCM follow their prescribed gallery-trial protocols, while RegDB follows its official ten-split protocol. Improvements are differences in percentage points.

Results and Interpretation. USIL⁺ improves both Rank-1 and mAP over USIL across all six retrieval settings, providing additional cross-dataset evidence for the contribution of explicit stripe-level reliability supervision. In particular, Rank-1/mAP improve by +3.04/+2.00 percentage points on SYSU-MM01 All-search, +3.01/+1.66 on RegDB IR→VIS, and +1.98/+1.98 on LLCM VIS→IR. These improvements also support the effectiveness of the finalized supervised configuration under the evaluated protocols. We restrict this conclusion to Rank-1 and mAP, as some SYSU-MM01 Rank-10 scores decrease relative to USIL.

Statistical Reliability and Limitations. The additional cross-dataset results and isolated ablations provide complementary evidence for the effectiveness of USIL and its components. However, the reported ten-gallery-trial and ten-split averages do not constitute independent training-seed evaluations; therefore, we do not claim cross-seed statistical significance.

Manuscript Revisions. We have revised Tables 1–2 to report both configurations and updated Appendix E.2 to clarify their training settings. We also clarify the evaluation protocols and the distinction between repeated evaluations and independent training runs in Secs. 4.2–4.3 and Appendix E.5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.