acceptodds
Under review as a conference paper at ICLR 2027

TRIDA: Tri-modal Residual-aware Inter-modal Divergence Arbitration for Multi-spectral Object Re-identification

Abstract

Multi-spectral object re-identification (ReID) aims to match identities across heterogeneous spectral observations. This matching capability is critical under varying illumination and adverse sensing conditions. Cross-spectral interaction is essential for exploiting complementary spectral cues, yet it is not cost-free: token selection discards potentially discriminative cues at unselected positions, all-token retention propagates modality-specific interference, and unconstrained interaction further incurs representation drift and gradient conflicts. We propose Tri-modal Residual-aware Inter-modal Divergence Arbitration (TRIDA), which addresses these challenges through three coordinated designs. Specifically, GRACE retains all tokens to preserve identity cues and reorganizes cross-modal information via channel-group cyclic permutation, while cross-modal CLS differences adaptively modulate CLS attention and residual scaling to mitigate modality-specific interference. GRASP then counters representation drift by regularizing group-wise deviations toward a sample-wise cross-modal center with progressive scheduling, while preserving discriminative modality-specific information. Finally, MCGA resolves conflicting gradients: it probes modality-wise gradients from the identity-discriminative and image-text objectives on the trainable GRACE enhancement parameters, and applies conflict-severity-adaptive, order-independent arbitration, leaving all unprobed gradient contributions intact. Through empirical evaluations on five multi-spectral object ReID benchmarks, TRIDA demonstrates superior object retrieval performance while requiring only modest parameter overhead.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-spectral object re-identification, Cross-modal interaction, Residual-aware representation, Gradient conflict arbitration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9HjDOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRIDA for RGB–NIR–TIR object re-identification, building on ICPL-ReID. It introduces GRACE for channel-group exchange and adaptive feature enhancement, GRASP for cross-modal feature consistency, and MCGA for mitigating conflicting modality gradients during training. The method is evaluated on five person and vehicle ReID benchmarks, with component ablations on RGBNT201 and RGBNT100.

Critical strengths

TRIDA combines cross-modal feature exchange, consistency regularization, and gradient conflict mitigation in a unified framework. It improves mAP over the ICPL-ReID baseline across five person and vehicle ReID benchmarks, supporting its effectiveness across different datasets and object categories.

Critical weaknesses

The experimental comparison omits recent relevant methods. There are some recent examples published at CVPR 2026: CoT-ReID (main conference), FusionBridge, and RoDI (Findings). Notably, CoT-ReID is cited in the paper but is not included in the comparison tables. These omissions make it difficult to assess whether TRIDA offers meaningful improvements over recent approaches.

Questions for authors
  1. The fixed channel exchange strategy is insufficiently justified. The paper does not explain why contiguous channel groups and the specific cyclic exchange pattern are appropriate, or examine alternative group numbers and exchange rules. Moreover, the GRACE ablation evaluates channel exchange together with the learnable enhancement components, leaving the contribution of the exchange operation itself unclear. Please explain why do you select this fixed channel exchange strategy.
  2. The paper does not provide sensitivity analyses for key hyperparameters, such as the number of channel groups, the GRASP regularization weight, and the MCGA conflict threshold. It therefore remains unclear how robust the reported gains are to these choices. Please provide hyperparameters sensitivity analyses data.
  3. MCGA computes the g_m using only the ReID and image-to-text losses. How would including the other loss such as image-to-prototype loss affect the detected conflicts and final retrieval performance? Could you provide a controlled comparison to assess the impact of this choice?
  4. The ablation results provide limited evidence for the effectiveness of MCGA. On RGBNT100, adding MCGA leaves all reported retrieval metrics unchanged. On RGBNT201, it improves mAP by only 0.3 percentage points, although Rank-1 increases by 1.6 points. Without repeated runs or results on additional datasets, it is unclear whether these gains are consistent and reproducible. Could the authors provide further evidence that MCGA offers a reliable benefit?
  5. On RGBNT201, comparing variants (c) and (d), adding MCGA improves Rank-1 by 1.6 percentage points, while mAP increases by only 0.3 points and Rank-5 and Rank-10 decrease slightly. I think it's weird. Could you explain this mixed pattern and report results over multiple runs to clarify whether the Rank-1 improvement is consistent?

I will improve the rating and confidence if the rebuttal provides clearer explanations and convincing supporting results.

Reviewer_22mbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework named TRIDA to address several challenges in multi-spectral object re-identification (ReID), including potential information loss, modality-specific interference, representation drift, and gradient conflicts arising from cross-modal feature interactions. Built upon ICPL-ReID, the proposed framework consists of three main modules. Specifically, GRACE facilitates cross-modal information exchange through channel-group cyclic permutation and adaptively refines feature representations using a cross-modal discrepancy-aware residual enhancement mechanism. GRASP introduces group-wise residual regularization to constrain the deviations of modality-specific features from a shared center, thereby improving cross-modal representation consistency. MCGA detects gradient conflicts among different modalities and adaptively adjusts the gradient update directions of shared parameters. The authors conduct experiments on five multi-spectral person and vehicle ReID benchmarks, complemented by ablation studies, feature visualizations, and gradient conflict analyses to evaluate the proposed approach. Experimental results show that TRIDA consistently improves mAP over the reproduced ICPL-ReID baseline across all five datasets and achieves competitive performance on several benchmarks.

Critical strengths
  1. The paper addresses relevant challenges in multi-spectral ReID, including information preservation, representation drift, and gradient conflicts. The motivation is clear and practically meaningful.
  2. The proposed framework integrates feature interaction, representation regularization, and gradient optimization, with each module serving a clearly defined purpose.
  3. Experiments on five person and vehicle ReID benchmarks, including comparisons with existing methods, ablation studies, and visualization analyses, provide empirical support for the proposed approach.
Critical weaknesses
  1. The three proposed modules mainly build upon existing techniques, including channel-group exchange, feature regularization, and gradient conflict resolution. Their designs are relatively straightforward, and the overall methodological novelty remains limited despite the complex formulation.
  2. Although GRACE preserves all tokens through channel-group exchange, it remains unclear how this strategy prevents irrelevant information propagation or cross-modal interference. More direct quantitative and qualitative evidence is needed to support its claimed advantages.
  3. The framework introduces multiple hyperparameters for loss balancing. However, their selection criteria and sensitivity are insufficiently explored, raising concerns about the robustness of the proposed method.
  4. TRIDA outperforms STMI on RGBNT201 but achieves lower mAP on the vehicle datasets MSVR310 and RGBNT100. The authors should investigate the reasons for these performance differences across person and vehicle ReID tasks.
  5. Figures 1 and 2 are overly crowded, with small text and unclear connections between components. Their layouts should be simplified and reorganized to improve readability and understanding of the overall framework.
Questions for authors
  1. The three proposed modules mainly build upon existing techniques, including channel-group exchange, feature regularization, and gradient conflict resolution. Their designs are relatively straightforward, and the overall methodological novelty remains limited despite the complex formulation.
  2. Although GRACE preserves all tokens through channel-group exchange, it remains unclear how this strategy prevents irrelevant information propagation or cross-modal interference. More direct quantitative and qualitative evidence is needed to support its claimed advantages.
  3. The framework introduces multiple hyperparameters for loss balancing. However, their selection criteria and sensitivity are insufficiently explored, raising concerns about the robustness of the proposed method.
  4. TRIDA outperforms STMI on RGBNT201 but achieves lower mAP on the vehicle datasets MSVR310 and RGBNT100. The authors should investigate the reasons for these performance differences across person and vehicle ReID tasks.
Reviewer_NcVAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TRIDA, a tri-modal framework for RGB-NIR-TIR person re-identification. It introduces group-wise feature exchange (GRACE) to enhance cross-modal interactions, a residual anchoring loss (GRASP) to improve representation consistency, and modality-aware gradient arbitration (MCGA) to mitigate conflicting gradients during training. Experiments on five benchmarks demonstrate the effectiveness of the proposed approach, although the ablation studies and efficiency analysis remain limited.

Critical strengths

Experiments on five benchmarks demonstrate the effectiveness of the proposed approach.

Critical weaknesses
  1. Insufficient Novelty: The novelty of the proposed framework appears limited, as it combines several established techniques, including cross-modal feature exchange, residual modulation, representation regularization, and gradient conflict resolution. The paper does not sufficiently clarify what fundamentally distinguishes GRACE, GRASP, and MCGA from existing methods or establish the conceptual novelty of their combination.

  2. Insufficient Ablation Studies and Poor Presentation Quality: The ablation studies are insufficient to fully validate the proposed design. They evaluate the modules primarily at the module level, without systematically examining the internal design choices and key factors within each module. For example, the individual components and hyperparameters of GRACE, GRASP, and MCGA are not sufficiently isolated to demonstrate their respective contributions. Moreover, the figures are poorly presented: many are excessively small, making their contents difficult or even impossible to read, while the overall figure quality is low. The visualizations are also largely qualitative and provide limited evidence for the claimed improvements in cross-modal consistency and interference suppression. The authors should substantially improve figure readability and presentation quality, and provide more fine-grained ablations to justify the proposed design choices.

  3. Efficiency and Scalability: Although the paper reports a 7.97% increase in trainable parameters, it provides limited analysis of computational overhead, training time, and memory consumption. Since MCGA requires additional modality-wise gradient computations during training, what is its actual computational cost, and does TRIDA offer a practical efficiency advantage over existing multi-spectral ReID methods?

Questions for authors

See Weaknesses.

1 comment

We thank the reviewer for the constructive assessment. We address novelty, fine-grained evidence, presentation, and computational overhead below with new experiments; the corresponding revisions are marked in red in the revised PDF.

W1. "Insufficient novelty: it combines several established techniques."

We acknowledge that feature exchange, regularization, and gradient conflict mitigation have individually been studied, and we do not claim these are new primitives. TRIDA's contribution is their specific coordination within an established ICPL-ReID framework, rather than a newly designed backbone: (i) GRACE simultaneously permutes channel groups across RGB/NIR/TIR without dropping token positions (two groups remain modality-native at G=4G=4), followed by discrepancy-conditioned enhancement with bounded adaptive residual scaling; (ii) GRASP softly penalizes group-wise deviations from a differentiable, sample-specific cross-modal center with warm-up, instead of enforcing identical modality representations; (iii) MCGA selectively arbitrates gradients on the shared enhancement parameters using cosine-thresholded, symmetric, order-independent corrections with severity-adaptive blending, retaining unprobed gradients and recovering the ordinary update when there is no conflict (Eq. (14)). These mechanisms address feature interaction, representation drift, and optimization conflicts at distinct stages.

W2. "Ablation studies are insufficient … figures are poorly presented; visualizations are largely qualitative."

We have added fine-grained analyses and improved the presentation:

  • Component-level ablation. Table 3 reports RGBNT201 mAP/Rank-1 of 73.0/74.8 (baseline), 80.6/81.5 (+GRACE), 82.3/83.7 (+GRASP), and 82.6/85.3 (+MCGA). New Table 4(a) decouples GRACE on WMVEID863: exchange only improves mAP/Rank-1 by +1.6/+2.1 points over the baseline (68.2/75.3 → 69.8/77.4), enhancement alone gives +0.1/+0.9, and full GRACE attains the best result (70.0/78.1). New Table 4(b) examines channel-group granularity on RGBNT201: G=2/3/4/6G=2/3/4/6 yields 78.2/79.9, 79.8/81.1, 80.6/81.5, and 78.4/80.1; G=4G=4 performs best among the evaluated configurations.
  • Repeatability. Three paired MCGA ON/OFF runs on RGBNT201 (seeds 1234/1235/1236) give mAP gains of +0.3/+0.5/+0.1 and Rank-1 gains of +1.6/+1.8/+0.9 points, with mean paired improvements of +0.30 ± 0.20 (mAP) and +1.43 ± 0.47 (Rank-1). All three seeds improve both metrics; we acknowledge that these runs do not establish statistical significance and that MCGA does not consistently improve mAP on every benchmark (e.g., RGBNT100).
  • Quantitative optimization evidence. New Appendix Fig. 8 plots gradient cosine similarities, conflict ratios, and effective arbitration strength: MSVR310 exhibits conflict ratios of approximately 0.3-0.5, whereas RGBNT201 shows weak conflict and negligible arbitration, confirming dataset-dependent, conflict-adaptive behavior.
  • Presentation. Fig. 1 is enlarged and Fig. 2 is simplified and reorganized. We acknowledge that the Grad-CAM evidence (Fig. 6) is qualitative, will improve figure readability, and now distinguish visualizations from direct interference measurements. Broader sensitivity analyses of GRASP and MCGA hyperparameters remain a stated limitation.

W3. "What is MCGA's actual computational cost, and does TRIDA offer a practical efficiency advantage?"

We benchmarked matched MCGA OFF/ON configurations on RGBNT201 (single RTX 4090, batch size 64; identical GRACE/GRASP/parameters/optimization, MCGA being the only difference; 3 runs, 10 warm-up + 40 synchronized timed iterations each):

Metricw/o MCGAw/ MCGAOverhead
Training time (ms/iter)268.395 ± 0.308280.516 ± 0.308+4.52%
Peak GPU memory (GiB)13.89813.950+0.38%

The overhead decomposes into 9.519 ms/iter for modality-wise gradient probing and 2.437 ms/iter for conflict arbitration. Both configurations contain 354.230M total and 49.036M trainable parameters, so MCGA adds no model parameters and no inference-time operations; TRIDA's inference cost equals the baseline plus GRACE alone. These measurements quantify training-step overhead (excluding data loading, prototype extraction, and evaluation). We do not claim a cross-method efficiency advantage based on these results; they show that the arbitration mechanism is lightweight relative to a training step.

We appreciate the reviewer's suggestions and have incorporated these substantiated analyses and clearer limitations into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.