acceptodds
Under review as a conference paper at ICLR 2027

Instance-Cluster Cross-Clothing Supervision for Sparse Unsupervised Clothing Change Person Re-Identification

Abstract

Clothing change person re-identification (CC-ReID) aims to match images of the same person captured under different clothing conditions. To reduce annotation costs, recent studies focus on unsupervised CC-ReID (UCC-ReID), which typically employs clustering and matching algorithms to generate identity pseudo-labels and associate images of the same person wearing different clothing. Although existing UCC-ReID methods substantially reduce the need for manual annotation, they still rely on a stringent dense identity assumption: the vast majority of individuals in the training set are assumed to have images captured in more than one outfit. However, this assumption is often difficult to satisfy in many public scenarios. To this end, we not only consider the dense scenario but also explore the more challenging sparse scenario, in which only a small proportion of identities in the training set have images captured in more than one outfit. We propose an instance-cluster cross-clothing supervision (ICCS) framework, which elegantly generates cross-clothing supervision signals at both the cluster and instance levels. At the instance level, ICCS introduces the random channel augmentation and random clothing fusion modules to transform the clothing regions of person images, thereby enhancing the model's robustness to clothing variations. At the cluster level, ICCS introduces an identity-clothing mutual supervision module, which enables the class vectors of original and clothing-masked images to mutually guide each other, thereby providing accurate cross-clothing supervision signals for optimization. Experimental results on multiple CC-ReID benchmarks under both dense and sparse scenarios demonstrate that the proposed ICCS significantly outperforms existing state-of-the-art methods in overall performance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
clothing change person re-identification, sparse identity, unsupervised learning, contrastive learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_4p8POpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies unsupervised clothing-change person re-identification when only a subset of training identities appears in multiple outfits. The proposed ICCS framework combines clothing-region color augmentation and clothing-texture fusion with mutual supervision between original and clothing-masked images. An online encoder is trained against momentum-encoder cluster centroids, while bidirectional cross-view KL losses align class-probability vectors. The approach avoids explicitly assigning a cross-clothing partner to every cluster. Experiments cover PRCC, LTCC, and VC-Clothes under the original training sets and two simulated sparse settings, with results averaged over five randomly generated sparse subsets.

Critical strengths
  1. The sparse setting exposes a relevant limitation, and the reported improvements are meaningful. Questioning cross-clothing matching when many identities lack a second observed outfit is well motivated. The paper evaluates multiple datasets and averages the sparse-setting results over five subset constructions. On PRCC(CC), ICCS reports 53.1% mAP under Sparse(0.25), compared with 45.6% for MALC; the advantages are also substantial on VC-Clothes. These results suggest practical value beyond the relatively small improvements in the dense setting.
Critical weaknesses
  1. The distinctive contribution and reliability of the proposed cross-clothing supervision are not sufficiently established. Clothing-specific augmentation already appears in CICL [1], and soft pseudo-label refinement through mutual teaching has an established precedent in MMT [2]. ICCS’s potentially distinctive contribution is therefore the use of aligned original/masked cluster distributions to provide supervision without forced cross-clothing matching, rather than augmentation or mutual teaching individually. These precedents do not duplicate ICCS, but stronger mechanism-specific evidence is needed. The central argument assumes that masked-image class vectors are bimodal when a cross-clothing positive cluster exists and unimodal otherwise. However, masked clusters inherit the original-image pseudo-labels, and the paper does not quantify whether secondary probability mass actually corresponds to the same identity rather than an unrelated person with similar non-clothing appearance. The distance plots examine positive pairs but do not establish the rate of incorrect cross-identity supervision. Moreover, Eq. (5) defines a centroid as an unnormalized sum; unless normalization is applied elsewhere, cluster size also affects the class-vector logits. The ablations demonstrate an aggregate gain from ICS, but do not isolate reciprocal supervision from one-way masked-to-original distillation or a matching baseline that can abstain from assigning a partner.

  2. The sparse-setting evidence contains an attribution confound and an inconsistent ablation description. The sparse datasets are constructed by removing entire clothing sets. Consequently, the intervention changes both cross-outfit availability and training-set size, and may also alter the camera/view distribution. Evaluating all methods on the same subsets remains a valid comparison, but the dense-to-sparse degradation cannot be attributed specifically to incorrect forced matching without a size-matched control and direct matching-error statistics. Averaging five subset constructions is useful, but does not separate these effects. There is also a concrete inconsistency in the main ablation setting. Section 4.3 states that Table 2 uses Sparse(0.50), but its numbers reproduce the Sparse(0.25) block of Appendix Table 3. For example, the full-model PRCC(CC) result is 53.1/43.2 in the main table, whereas the appendix reports 55.3/43.7 for Sparse(0.50). This may be a labeling error, but it must be corrected because the accompanying component analysis is presented as evidence for a specific sparsity regime.

[1] Pang, Zhao, and Wang. Clothing-Invariant Contrastive Learning for Unsupervised Person Re-Identification. Neural Networks, 2024.

[2] Ge, Chen, and Li. Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification. ICLR, 2020.

Questions for authors
  1. Can you quantify how much masked-teacher probability mass is assigned to true cross-outfit positives versus unrelated identities, separately for identities with and without another observed outfit? Does bidirectional ICS outperform one-way masked-to-original supervision under otherwise identical settings?

  2. Are the features and centroids normalized before Eqs. (6)–(8)? Please reconcile the unnormalized sum in Eq. (5) with the interpretation of the resulting class distributions, and specify the clustering distance and DBSCAN radius used.

  3. Which sparsity setting was actually used for the main Table 2 and its associated analyses? Can you provide retained image/camera counts and a size-matched control to distinguish the effects of reduced data volume from reduced cross-clothing availability?

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

This manuscript focuses on the field of person re-identification and sets up a sparse unsupervised clothing change person re-identification task. The proposed task has practical significance. For this task, the authors proposed an instance-cluster cross-clothing supervision (ICCS) method. ICCS first enables the class vectors of real images and masked images to mutually guide each other, Afterwards, data augmentation methods are introduced to encourage encoders to learn clothing-invariant and identity-relevance features. The experimental results on multiple datasets have validated the superiority and generalization of the proposed ICCS.

Critical strengths
  1. This manuscript innovatively proposes sparse UCC-ReID, which is a task that meets the needs of real-world applications.
  2. The authors have designed a novel and effective method for the proposed new task, which is inspiring for future research.
Critical weaknesses

None

Questions for authors
  1. Although the proposed task and method are meaningful, there are some concerns about the flexibility of the proposed ICCS. Specifically, ICCS relies on an additional human parsing model. Does it take a lot of time to choose a good human parsing model?
  2. Are RCA and RCF offline or online? Specifically, the authors should state whether the two proposed data augmentation methods will significantly increase training time?
1 comment

The authors' responses regarding the human parsing model and data augmentation have effectively addressed my concerns. Overall, the contributions of this paper mainly focuses on a meaningful new task rather than offering a technical breakthrough. I have no further concerns and will raise my score to 4.

Reviewer_Ly9LOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Instance-Cluster Cross-Clothing Supervision (ICCS) for unsupervised clothes-changing person Re-ID in both dense and sparse scenarios. To address the noise introduced by forced cross-clothing cluster matching in sparse scenarios, ICCS generates cross-clothing supervision at both instance and cluster levels: at the instance level, Random Channel Augmentation (RCA) and Random Clothing Fusion (RCF) alter clothing color and texture, respectively; at the cluster level, Identity-Clothing Mutual Supervision (ICS) enables the class vectors of real images and clothing-masked images to mutually guide each other. The model is optimized with online and momentum encoders and an identity-centroid contrastive loss. It outperforms existing methods on multiple datasets

Critical strengths
  • The paper addresses sparse unsupervised clothes-changing Re-ID, a practical but underexplored setting, and reveals the limitations of cross-clothing cluster matching when few identities have multiple outfits.
  • Extensive experiments on three benchmarks under dense and sparse settings, including comparisons, ablations, and feature visualizations, demonstrate effectiveness, especially when cross-clothing positives are limited.
Critical weaknesses
  • The methodological novelty of the proposed framework is limited. RCA mainly combines conventional hue perturbation and RGB channel permutation, while RCF performs region-level clothing mixing, both of which are closely related to existing augmentation techniques. The ICS module employs bidirectional KL divergence to align class distributions between original and clothing-masked images, resembling established cross-view consistency learning and knowledge distillation strategies. In particular, previous methods such as CICL and JAPL have already explored clothing-specific augmentation and class-vector consistency for unsupervised CC-ReID. Although the proposed modules achieve performance improvements, the paper does not sufficiently clarify their fundamental differences from these existing approaches. Consequently, the overall framework appears to be largely an incremental combination of established techniques rather than a substantial methodological advancement.
  • The experimental evidence is insufficient to fully support the claimed advantages of ICS in sparse scenarios. In Section 3.4, the authors argue that clothing-masked class vectors exhibit different distribution patterns depending on the existence of cross-clothing positive clusters, allowing ICS to avoid incorrect supervision caused by forced matching. However, this assumption is not directly validated through quantitative analysis. Since the proposed loss mainly enforces consistency between original and masked views, it remains unclear whether ICS can reliably establish cross-clothing relationships or simply improve cross-view feature consistency. Furthermore, the sparse settings in Appendix B.1 are constructed by removing clothing subsets, which simultaneously reduces clothing diversity and the amount of training data. The absence of data-volume-controlled experiments and comparisons with confidence-based cluster matching methods makes it difficult to determine whether the reported improvements specifically arise from addressing sparse cross-clothing supervision.
Questions for authors
  1. Since the sparse settings are constructed by removing clothing subsets, can the authors conduct experiments with a controlled total number of training images to separate the effects of clothing sparsity from reduced training data? Additionally, can the authors compare ICCS with a confidence-threshold-based cluster-matching baseline to verify whether the proposed mechanism offers advantages beyond simply avoiding unreliable matches?
  2. Section 3.4 suggests that the class vectors of clothing-masked images exhibit different distribution patterns depending on whether cross-clothing positive clusters exist. Can the authors provide quantitative evidence supporting this assumption and explain how ICS reliably establishes supervision between different clothing clusters belonging to the same identity?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.