acceptodds
Under review as a conference paper at ICLR 2027

Mapping and Matching Joint Learning for Unpaired Unsupervised Visible-Infrared Person Re-Identification

Abstract

Unsupervised visible-infrared person re-identification (UVI-ReID) aims to associate visible and infrared images of the same identities without relying on manual annotations. Existing approaches typically rely on the unrealistic assumption that the identities between the two modalities are completely overlapped. To advance the development of UVI-ReID, recent studies begin to explore the unpaired UVI-ReID task and introduce feature mapping strategies. Although these studies achieve certain progress, they overlook two critical factors: the identity consistency of the mapped features and the effectiveness of inter-modality overlapped identities (i.e., real cross-modality positive pairs). To this end, we propose a mapping and matching joint learning method (MMJL). First, we introduce an attention-guided feature mapping module, which reconstructs and weights features based on cross-modality attention weights to generate mapped features while preserving intra-identity consistency. Then, we introduce a cross-modality positive matching module, which identifies real cross-modality positive clusters based on the consistency between two cross-modality matching results and the similarity between the intra-modality mapped features and the target features. Moreover, we introduce multiple contrastive losses to simultaneously ensure the identity discriminability and modality invariance of person features. Extensive experiments on two unpaired visible-infrared datasets demonstrate the superiority of the proposed method and the effectiveness of each component.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
visible-infrared person re-identification, unpaired identity, unsupervised learning, contrastive learning

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_nmDvOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the unpaired unsupervised visible-infrared person re-identification (UVI-ReID) task, where the identity sets of the two modalities are only partially overlapped. The authors propose MMJL, a joint mapping and matching learning framework consisting of three key components: an attention-guided feature mapping module (AFM), a cross-modality positive matching module (CPM), and an inter-modality discriminative contrastive loss. Extensive experiments on unpaired SYSU-MM01 and RegDB under overlap ratios of 0.25, 0.50, and 0.75 show consistent improvements over prior methods, with particularly large gains in low-overlap regimes. Ablation studies validate the contribution of each component.

Critical strengths
  1. The paper identifies two genuine limitations of the existing MCL baseline and addresses each with a technically sound module. The two-stage design that explicitly separates generated positive pairs from mined real positive pairs is a principled way to diversify cross-modality supervision.
  2. The method is evaluated across three overlap ratios on two datasets, and the reported gains are consistent across all settings, which is rare in this subfield. The analysis showing that the advantage grows as the overlap ratio decreases directly supports the paper's core claim that exploiting overlapped identities is especially valuable in highly unpaired scenarios.
Critical weaknesses

None

Questions for authors
  1. The cross-modality positive clusters are identified based on the consistency of two matching results plus a similarity threshold h. Since both matching results are derived from features produced by the same imperfect encoder, could confirmation bias arise where early matching errors are reinforced and propagated across epochs?
  2. The final mapped feature in Eq. (3) fuses the reconstructed feature with the original feature using a residual term weighted. Could the authors clarify why this specific weighting (rather than, e.g., a learnable gate or a fixed ratio) is appropriate?
Reviewer_yR4xOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript focuses on unpaired unsupervised visible-infrared person re-identification. The tasks targeted have certain practical basis. The authors introduced three interrelated components. AFM considers the inter-modality identity similarity relationships during the feature mapping process, thereby preserving the intra-identity consistency. CPM leverages the consistency of cross-modality matching results and the similarity of intra-modality features to facilitate modality-invariant feature learning. The introduced loss functions can utilize generated and real cross-modality positive pairs to promote model optimization. The authors evaluate the performance under three different settings on each dataset, validating the effectiveness of the proposed MMJL.

Critical strengths
  1. The proposed method is novel and reasonable, and each component can be well connected together to enhance effective handling of modality gaps and identity mismatches.
  2. The authors not only conduct experiments on three settings, but also reproduced existing methods, providing a comparable reference for subsequent research. The experimental results also indicate that the proposed method has superiority.
Critical weaknesses

None

Questions for authors
  1. The authors claim that the proposed AFM can maintain intra-identity consistency and provide corresponding experimental results. However, the authors should answer what is the theoretical basis for AFM to maintain intra-identity consistency.
  2. The authors set two evaluation criteria in CPM to find highly similar cross-modality positive clusters. Although these two evaluation criteria are logical, the authors should also provide the reasons for such design.
Reviewer_pNciOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper targets unpaired visible-infrared person re-identification, where the identity sets of the two modalities are only partially overlapped. It argues that MCL’s mean and variance-based feature mapping weakens intra-identity consistency and ignores real cross-modality overlapped identities. The authors propose MMJL, which combines an attention-guided feature mapping module, a cross-modality positive matching module, and several contrastive losses. Experiments on unpaired SYSU-MM01 and RegDB under 0.25, 0.50, and 0.75 overlap ratios show consistent gains over recent methods. Ablation studies support the individual contributions of AFM, CPM, and the inter-modality discriminative contrastive loss.

Critical strengths
  1. The AFM design is a clean response to a concrete weakness in MCL, instead of relying only on modality-wise mean and variance. It reconstructs features from target-modality centroids and then fuses them with the original features. The visualizations in Fig. 3 and the ablation results give credible evidence that this improves the compactness of mapped features with same-identity target features. The module is simple and easy to plug into the iterative clustering pipeline.

  2. The CPM module addresses a real gap in unpaired UVI-ReID by trying to recover real cross-modality positive pairs from partially overlapped identities. Using both consistent graph-matching results and a similarity-based fallback for inconsistent matches is a practical way to enrich supervision. The performance advantage at the 0.25 overlap ratio is particularly encouraging.

Critical weaknesses

None

Questions for authors
  1. The fallback selection in CPM uses a top-h ratio over cosine similarities. How sensitive is this step to model performance? Could the same identity be incorrectly added as a positive cluster under such conditions?

  2. The compared methods are reproduced under a newly defined unpaired setting. Could you clarify the exact training/testing split protocol and whether MCL and BMIL were tuned with the same overlap ratios, batch composition, and clustering hyperparameters?

1 comment

I thank the authors for their responses regarding CPM, the unpaired setting, and the experimental details. I have no further concerns and have therefore raised my score to 4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.