- 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
- 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
-
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.
-
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
-
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?
-
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.
-
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?
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.