acceptodds
Under review as a conference paper at ICLR 2027

Asymmetric Domain Knowledge Distillation for Unsupervised Cross-Domain Person Re-Identification

Abstract

Unsupervised domain‑adaptive person re‑identification (UDA‑ReID) is plagued by noisy clustering‑derived pseudo‑labels, inherent feature extraction errors from cross‑domain style discrepancies, underutilized outliers, and ineffective source‑domain knowledge transfer in target fine‑tuning. Existing methods either focus on clustering or pseudo‑label refinement while ignoring cross‑domain feature bias, or use contrastive learning hindered by unreliable clustering and marginal value of cross‑domain negative samples. To address these issues in this paper, UDA‑ReID is formulated as an asymmetric domain distillation task, distinct from traditional single‑domain teacher‑student distillation by transferring identity‑related feature clustering from the labeled pre‑trained source to unlabeled target domain. A novel asymmetric domain distillation method is proposed, where identity and domain‑related features are decoupled and recoupled to mitigate domain gaps. And a graph network‑based feature transfer structure is designed for cross‑domain transformation, along with a multi‑domain pseudo‑label refinement strategy leveraging complementary clustering information to generate accurate and noise‑robust target pseudo‑labels. Extensive experiments validate the effectiveness of the proposed framework.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Person re-identification, Unsupervised domain adaptation, Graph convolutional network, Complementary clustering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KH1TOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes ADKD for unsupervised domain-adaptive person re-identification. It uses domain-specific batch normalization, a learned identity-agent filter, and domain-confusion training to construct identity-related representations. A graph network combines these representations with domain-style nodes to produce alternative domain embeddings. The resulting clusterings are used for pseudo-label assignment and sample selection, based on a combination of clustering quality, changes in pairwise cluster assignments, and cross-domain agreement. Experiments report single-source and multi-source adaptation results, component ablations, and feature visualizations.

Critical strengths
  1. Exploiting complementary clustering spaces is a relevant research direction. Rather than relying exclusively on clustering in the target model’s native embedding, the method attempts to use alternative domain representations to improve pseudo-label assignment and sample selection. The component ablations report incremental improvements, providing an initial empirical rationale for investigating this design. However, the formulation and reporting problems below prevent these results from establishing the claimed contribution convincingly.
Critical weaknesses
  1. The methodological novelty over existing multi-domain adaptation frameworks is insufficiently established. RDSBN+MDIF [1] already combines domain-specific normalization, learnable weighted domain-agent nodes updated through a moving average, and GCN-based cross-domain information fusion. ADKD’s domain-style construction in Eqs. (5)–(6) and its graph-based processing closely resemble these established components. Although identity filtering and adaptive clustering-space selection introduce additional elements, the manuscript does not sufficiently establish their distinctive contribution beyond extending an existing domain-fusion framework. The cumulative ablations demonstrate gains from adding modules, but do not isolate the proposed transformation from alternative fusion mechanisms. A controlled comparison replacing DT-GCN with MDIF while retaining the same filtering and pseudo-label refinement would clarify this incremental contribution.

  2. Dataset and protocol inconsistencies undermine the main experimental comparisons. Table 1 lists 38.0/68.1/79.0/81.7 for DKD-MPL under Market→Duke, but this tuple matches its published MSMT17 result. The original paper instead reports 75.9/86.8/92.8/94.0 on DukeMTMC-reID, under a purely unsupervised setting. The comparison therefore requires correction of both dataset attribution and supervision-protocol labeling; the published purely unsupervised results should not be presented as unspecified source-to-target UDA runs. Table 2’s first two columns contain the same source datasets and target dataset, merely reordered, with identical results throughout. Unless an order-dependent protocol is intended and explained, these are duplicate settings rather than independent evaluations. These reporting issues materially weaken the claimed superiority and warrant an audit of the comparison tables。

[1] Zechen Bai, Zhigang Wang, Jian Wang, Di Hu, and Errui Ding. Unsupervised Multi-Source Domain Adaptation for Person Re-Identification. CVPR, 2021.

[2] Wenjie Zhu, Bo Peng, and Wei Qi Yan. Dual Knowledge Distillation on Multiview Pseudo Labels for Unsupervised Person Re-Identification. IEEE Transactions on Multimedia, 2024.

Questions for authors
  1. What exact expressions are implemented for Eqs. (7), (9), and (10)? Please resolve the silhouette sign, the inter-cluster averaging denominator, and the handling of negative or zero-sum confidence weights. Do the reported results use the printed equations or a different implementation?

  2. Please provide the complete training objective and DT-GCN update equations, including how destination-domain transformation is supervised, which parameters each loss updates, how unclustered outliers are labeled, and which representation is used at inference. These details are necessary to assess the proposed algorithm rather than optional implementation refinements.

  3. Please audit Table 1 against the original sources, particularly the DKD-MPL and IDE rows, and distinguish published purely unsupervised results from reproduced UDA results. Also explain whether Table 2’s first two columns are intentionally the same task. Do the main superiority claims remain valid after correcting the dataset and protocol assignments?

Reviewer_qYNnOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies unsupervised domain adaptive person re-identification using labeled source data and unlabeled target data. The authors argue that noisy target pseudo-labels arise not only from clustering errors but also from biased cross-domain features. ADKD reconstructs features through shared identity agents and a domain-confusion objective, then combines the reconstructed features with domain nodes in a DT-GCN to represent target samples in multiple domain spaces. It selects a space for pseudo-label assignment and gradually selects training samples using agreement among the resulting cluster relations. The paper reports improvements on single-source and multi-source transfer tasks, supported by component ablations and visualizations.

Critical strengths

The paper identifies a consequential limitation of cluster-based UDA-ReID: cross-domain feature bias may constrain the quality of any downstream clustering or label refinement. This provides a meaningful motivation for combining feature transformation with pseudo-label selection. The reported single-source transfer results are competitive, and the cumulative ablation in Table 3 shows gains as the main components are added. These results make the system worth examining further.

Critical weaknesses
  1. The central space-selection rule is not internally clear as written. Equation 7 defines \psi as mean within-cluster distance and \phi as mean distance to the nearest other cluster, but uses ((\psi-\phi)/\max(\psi,\phi)). Better separated clusters thus receive a more negative score, whereas Equation 9 maximizes its product with the nonnegative relation-change term \beta. The denominator used for the other-cluster mean in Equation 7 also needs checking. If implemented literally, selection may run counter to the stated motivation; if this is a notation error, the manuscript does not establish which algorithm produced the results.
  2. Moreover, \beta measures changes in pairwise cluster assignments, not their correctness. The paper does not quantitatively show that the selected space produces more accurate pseudo-labels or that erroneous selections are not reinforced during self-training. The claim that FL removes domain information while retaining identity information also lacks direct measurements of both properties. In the single-source setting, target samples can still be clustered in source and target spaces, but the paper does not show that dynamic selection improves over consistently using either one.
  3. Serious errors in the comparison tables undermine the claimed performance advantage. The original DKD-MPL paper explicitly studies purely unsupervised training without labeled source data. Its Table II reports 86.0 / 94.5 / 98.0 / 98.5 on Market-1501; the submission lists the same numbers under Duke→Market. Its Table IV reports 38.0 / 68.1 / 79.0 / 81.7 on MSMT17; the submission lists them under Market→Duke and again under Duke→MSMT17. The Market→Duke entry therefore does not even match the target dataset of the original result. All eight metrics in the IIDS and DMD rows of Table 1 are identical and require verification. The first two columns of Table 2 contain the same three-source-to-MSMT17 setting in a different order, with identical results, so they do not constitute two independent experiments. Table 2 also lacks recent multi-source UDA ReID baselines: its latest listed baseline dates from 2021, despite relevant work published in 2024 and 2025. The “DT-GAN” entry in Table 3(a) and the figure and table legibility issues add to the difficulty of checking the work. Most consequentially, the “outperforms existing methods” claims cannot be evaluated from Table 1 as currently presented.
Questions for authors
  1. Are the sign of the silhouette expression and the other-cluster denominator in Equation 7 correct? Do the implemented calculations in Equations 9–10 match the manuscript? Please identify which implementation produced all reported results.
  2. Can the authors quantify domain predictability and identity discrimination before and after FL, and report pseudo-label quality in each domain space? These results would directly test the proposed disentanglement and space-selection claims.
  3. In Duke→Market and similar tasks, which of the source and target spaces is selected at each epoch? How do dynamic selection, fixed-source selection, and fixed-target selection compare in pseudo-label quality and final performance?
  4. Is one of the first two Table 2 headers incorrect, or are the columns intentionally the same setting? Have the identical IIDS and DMD results in Table 1 been verified? Please state the number of distinct evaluation settings and clarify protocol differences from more recent multi-source methods.
Reviewer_W1G3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of unsupervised domain-adaptive pedestrian re-identification (UDA-ReID) and proposes an asymmetric domain knowledge distillation method. The core challenges of UDA-ReID include: pseudo-label noise arising from clustering; feature-extraction errors caused by cross-domain style differences; underutilization of outlying samples; and the low efficiency of transferring source-domain knowledge during fine-tuning on the target domain. The authors formulate UDA-ReID as an asymmetric domain distillation task, whereby knowledge derived from identity-related feature clustering on the pre-trained source domain is transferred to the unlabeled target domain. Experiments were conducted on four standard datasets using both single-domain and multi-domain adaptation approaches, and the authors report satisfactory results.

Critical strengths

(1) The agent-based attention-based feature decomposition mechanism in the Filter Layer avoids an explicit domain alignment loss function, resulting in a concise and effective design. The approach of explicitly decoupling identity features from domain features and then re-coupling them aligns with the prevailing trend toward feature decoupling; however, the authors' implementation of cross-domain transformation via a graph network provides a unique contribution. (2) The MLA module evaluates label quality by utilizing the contour coefficient to simultaneously measure intra-cluster compactness and inter-cluster separation, making it more comprehensive than simple clustering confidence threshold methods. The MSS module filters reliable samples through a cross-domain consistency check (XOR operation), effectively leveraging the complementarity of multi-domain clustering spaces. (3) The authors conducted thorough experiments under both single-domain and multi-domain settings and compared their results with those of more than nine recent methods. In particular, experiments on the large-scale MSMT dataset outperformed the second-best method by 2.6 percentage points, and achieved an absolute improvement of up to 8.8% in mAP under multi-domain settings, providing strong evidence for their effectiveness. (4) The paper is well-structured, with a logical flow from problem analysis to method design and experimental validation. The comparative diagram in Figure 1 (clustering methods vs. contrastive learning vs. traditional distillation vs. domain distillation) intuitively and clearly conveys the motivation behind the proposed method.

Critical weaknesses

(1) The paper emphasizes that its domain distillation is "asymmetric," distinguishing it from traditional synchronous point-to-point feature transfer. However, the paper's formal definition of "asymmetry" is not sufficiently clear: does it refer to the misalignment of the feature spaces of the source and target domains? Or does it refer to a unidirectional distillation direction (source → target)? Or does it refer to the transfer of clustering knowledge rather than the transfer of feature values? (2) The paper selects Graph Convolutional Networks (GCN) as the core architecture for domain transformation, but provides insufficient justification for this choice. Why is GCN more suitable for this task than MLPs, Transformers, or other nonlinear mappings? How is the graph structure determined? There is a lack of sensitivity analysis regarding the impact of these design choices on performance. Furthermore, the domain node update strategy in DT-GCN differs fundamentally from the traditional Exponential Moving Average (EMA); a more in-depth discussion is warranted. (3) ADKD comprises multiple components, each introducing additional computational overhead. The paper does not provide efficiency metrics such as training/inference time, parameter count increase, or FLOPs. Given the efficiency requirements in practical UDA-ReID deployments, the lack of an efficiency comparison with baseline methods is a significant shortcoming. In particular, the computational cost of graph construction and message passing in DT-GCN on large-scale datasets needs to be evaluated. (4) The paper notes that existing methods that directly discard outliers lose valuable information, and proposes leveraging outliers through multi-domain complementarity. However, the evaluation based on the contour coefficient in MLA inherently still tends to favor samples that are "easy to cluster"; truly difficult outlying samples (such as occluded or blurred images) may still be filtered out. The paper does not provide a quantitative analysis of the effectiveness of using outlying samples to validate this claim. (5) The symbol definitions in some formulas are not sufficiently consistent; for example, there is no clear explanation linking the similarity calculation in Equation (1) with the distance metric ρ() for the contour coefficient in Equation (7). In Equation (4), the formulation of the domain confusion loss contains indexing confusion, as Nd is used both to denote the number of domains and as the summation index. (6) The paper does not provide an analysis of cases in which the method fails. Error analysis can help identify the scenarios in which the method performs poorly (e.g., for specific domain pairs, specific attribute changes, or specific image qualities), which is highly valuable for assessing the method's practical limitations and guiding future improvements.

Questions for authors

Q1: How is the number of agents in the Filter Layer in the paper determined? What is the initialization strategy for the agents (random initialization, K-Means cluster centers, or another approach)? How does the number of agents affect final performance? Has sensitivity analysis been conducted regarding the number of agents? Q2: How are edges between domain nodes defined in DT-GCN? Is the graph fully connected, or is it a sparse graph constructed based on a certain similarity threshold? To what extent does the graph structure influence the effectiveness of cross-domain feature transfer? Q3: In the multi-domain pseudo-label refinement strategy, is there a conflict between the joint optimization of MLA and MSS? For example, MLA tends to select samples with high silhouette coefficients (i.e., samples that are easy to cluster), whereas MSS employs cross-domain consistency filtering; could this lead to a conflict in sample selection? Was such a conflict observed in the paper, and if so, how was it addressed?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.