acceptodds
Withdrawn from ICLR 2027

Interactive Semantic Verification for Unsupervised Domain Adaptation in Text-to-Image Person Re-Identification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unsupervised domain adaptation (UDA) has received increasing attention in text-to-image person re-identification (TIReID), as it enables models to adapt to an unlabeled target domain without costly annotations. However, under domain shift, existing feature-based adaptation methods face reliability challenges in both target domain pseudo correspondence construction and cross-domain representation learning. To this end, we propose an Interactive Semantic Verification framework (ISV) that leverages a multimodal large language model (MLLM). Specifically, to alleviate noisy pseudo image-text correspondences in the target domain caused by unreliable feature similarity, we develop Interactive Multi-evidence Learning (IML). It first performs balanced reciprocal matching to identify candidate correspondences and then employs an MLLM for semantic verification, leveraging external semantic knowledge to compensate for unreliable feature evidence, thereby removing mismatched pairs and recovering missed positive matches. Moreover, to improve cross-domain representation learning, we introduce Density-filtered Intermediate-domain Learning (DIL), which identifies and filters extreme feature outliers in low-density regions and explicitly models the difference between source and target features to construct more reliable intermediate representations. Extensive experiments on three TIReID benchmarks demonstrate that ISV consistently improves cross-domain retrieval performance and generalizes well across different source models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text-to-Image Person Re-Identification, Unsupervised Domain Adaptation, Cross-Modal Retrieval, Multimodal Large Language Models, Interactive Semantic Verification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YGVPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes interactive semantic validation (ISV), which is an unsupervised domain adaptation for text-to-image personnel re-identification. ISV contains three components: interactive multi-evidence learning (IML), which combines peer feature matching with MLLM-based semantic verification to improve the target pseudo-image-text correspondence; density filtering intermediate domain learning (DIL), which filters out-of-group values of low-density features and constructs intermediate domain representations. ; and text-guided spatial gate control (TSG), which guides the source text to the image-only space gate. Experiments on CUHK-PEDES, ICFG-PEDES and RSTPReid showed improvements in multiple transmission directions and source models.

Critical strengths
  1. ISV addresses both unreliable target pseudo-correspondences and cross-domain representation learning.
  2. IML uses a frozen MLLM to complement unreliable feature similarity and supports both false-pair filtering and missed-pair recovery.
  3. The paper evaluates multiple datasets, transfer directions, source backbones, MLLM scales, and reports three-seed stability, prompts, failure cases, and computational cost.
Critical weaknesses
  1. The DIL objective has a potential degeneracy. Since fm=λsfs+λtftf_m=\lambda_s f_s+\lambda_t f_t, the proposed feature-bridge loss simplifies to a term proportional to λsλt∥fs−ft∥2\lambda_s\lambda_t\|f_s-f_t\|^2, which is minimized when the mixing weight collapses toward either source or target rather than maintaining an intermediate representation. This appears inconsistent with the stated motivation of DIL.
  2. The target-domain protocol requires further clarification. Although the target training set is stated to be unlabeled and unpaired, the paper does not clearly explain whether target images and texts are independently shuffled and sampled during adaptation. This is important because preserving the original image-caption pairing in minibatches would leak target correspondence information.
  3. The ablation study does not sufficiently isolate the proposed components. The current progressive ablation cannot independently establish the contributions of DIL and TSG, and IML lacks filter-only/recovery-only and multi-round verification ablations. Some incremental improvements are also relatively small compared with the reported run-to-run variation.
  4. The comparison could be stronger and more resource-matched. Recent directly related UDA methods discussed in the paper, such as DMA, are absent from the main comparison, while ISV additionally benefits from a strong Qwen3-VL verifier. A simpler MLLM-assisted or resource-matched baseline would better establish whether the gains come from the proposed verification strategy rather than additional external semantic knowledge.
  5. Some methodological details remain insufficiently validated. The paper does not report quantitative precision/recall for MLLM filtering and recovery, and several TSG regularization terms such as LareaL_{\mathrm{area}}, LsmoothL_{\mathrm{smooth}}, and LbinaryL_{\mathrm{binary}} are described but not explicitly defined mathematically.
Questions for authors
  1. Are target images and target texts independently shuffled/sampled during adaptation, such that the original target image-caption correspondences are never accessible to the training pipeline?
  2. For DIL, since the feature-bridge objective can reduce toward zero when the mixing weights collapse to one domain endpoint, what prevents λs\lambda_s or λt\lambda_t from degenerating toward 1? Could the authors report the learned mixing-weight distributions?
  3. Could the authors clarify the individual contributions of false-pair filtering and missed-pair recovery in IML, e.g., by reporting filter-only and recovery-only results?
  4. Why are recent directly related UDA methods such as DMA not included in the main comparison, and how does ISV compare with a simpler MLLM-assisted baseline using a similar external-model budget?
Reviewer_ifaxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Interactive Semantic Verification (ISV), a framework for unsupervised domain adaptation in text-to-image person re-identification. The framework consists of three components: Interactive Multi-evidence Learning (IML), which employs MLLM-based semantic verification to refine pseudo image-text correspondences; Density-filtered Intermediate-domain Learning (DIL), which filters unreliable features and constructs intermediate-domain representations; and Text-Guided Spatial Gating (TSG), which transfers textual guidance to visual representations. Experiments on three benchmarks demonstrate competitive cross-domain retrieval performance across multiple transfer settings and source models.

Critical strengths
  1. The paper addresses an important and practical challenge in unsupervised domain adaptation for text-to-image person re-identification, particularly the reliability of pseudo image-text correspondences under domain shift.

  2. The proposed framework combines complementary strategies for pseudo-pair refinement, cross-domain feature learning, and spatial feature enhancement. The overall design is reasonably motivated and clearly presented.

  3. The experimental results demonstrate competitive performance across multiple transfer settings. The method also achieves consistent improvements when applied to different source models, suggesting reasonable generalizability.

Critical weaknesses
  1. The methodological novelty is relatively limited, as the framework mainly integrates and modifies existing techniques without introducing a sufficiently distinctive learning mechanism.

  2. The DIL bridge loss may encourage degenerate mixing weights, raising concerns about whether it effectively produces meaningful intermediate-domain representations.

  3. The improvements contributed by individual modules are relatively modest, and the current ablation studies do not sufficiently establish their effectiveness and necessity.

  4. The experimental evaluation lacks critical comparisons to justify the multi-round MLLM verification strategy, particularly the separate contributions of false-pair filtering and missed-pair recovery.

  5. The target-domain sampling protocol is not sufficiently clarified, and comparisons with some relevant existing methods remain incomplete.

Questions for authors
  1. Substituting Eq. (10) into Eq. (29), the feature bridge loss can be simplified to 2λsλt∥fs−ft∥22\lambda_s\lambda_t\|f_s-f_t\|_2. When the source and target features are fixed, minimizing this loss may drive the mixing weights toward extreme values rather than producing meaningful intermediate-domain representations. The authors should clarify how this potential degeneracy is avoided and provide an analysis of the learned mixing weight distributions, along with comparisons against fixed-weight feature mixing.

  2. The proposed method constructs pseudo image-text correspondences through batch-level feature matching in an unlabeled target domain. However, the sampling strategy for target images and texts is not sufficiently clarified. The authors should specify whether the two modalities are sampled independently, whether the original image-text correspondences are used during batch construction, and how they ensure that no ground-truth target correspondence information is exploited during training.

  3. According to Table 4, the complete framework improves Rank-1 accuracy by only 1.80 and 0.95 percentage points over the baseline on R→C and C→R, respectively. Moreover, introducing IML alone slightly decreases Rank-1 accuracy on C→R. The authors should clarify the exact configuration of the baseline used in the ablation study and report results over multiple random seeds for key component combinations to better substantiate the effectiveness of each module.

  4. Although multi-round MLLM-based semantic verification is a central contribution, the current experiments do not sufficiently demonstrate the necessity of the three-round interaction strategy. The authors should compare single-round verification, attribute extraction with rule-based decisions, and the complete three-round verification process. In addition, the accuracy, individual performance contributions, and computational costs of false-pair filtering and missed-pair recovery should be evaluated separately.

  5. The proposed framework mainly combines existing techniques, including feature-based matching, MLLM-assisted semantic verification, outlier filtering, and teacher–student distillation. Its methodological novelty over existing approaches remains insufficiently established. The authors should better clarify the key differences from prior MLLM-assisted TIReID methods and include experimental comparisons with relevant approaches such as DMA, which is already discussed in the paper, or explain why such comparisons are not feasible.

Reviewer_dgF3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies unsupervised domain adaptation for text-to-image person re-identification. It addresses incorrect and missed pseudo image-text pairs in the target domain under domain shift. The authors propose an Interactive Semantic Verification (ISV) framework. IML first uses balanced reciprocal matching to find candidate image-text pairs. A frozen multimodal large language model (MLLM) then checks fine-grained semantic conflicts, removes mismatched pairs, and attempts to recover missed positive pairs. DIL uses density estimates to filter extreme feature outliers. It also predicts mixing weights from differences between source and target features. TSG uses paired source text to guide spatial gating, then distills this guidance into an image-only student branch. The method is evaluated on CUHK-PEDES, ICFG-PEDES, and RSTPReid. The authors also test it with different source models. The results are competitive overall. However, the target-domain pairing protocol and the DIL objective still need clarification.

Critical strengths
  1. Rather than relying only on feature similarity to generate pseudo supervision, the paper uses an MLLM to provide additional semantic evidence. IML not only filters mismatched pairs but also attempts to recover positive pairs missed because of feature errors. This helps improve both the accuracy and coverage of target-domain pseudo supervision.

  2. The paper evaluates cross-domain retrieval on three public datasets and reports competitive results overall. Applying ISV to IRRA, CADA, and RDE improves performance in both C→I and I→C. This suggests that the framework is not tied to one source model. The multi-seed results in the appendix also provide evidence of its stabili

Critical weaknesses
  1. Section 3.1 states that the target domain has no original image-text correspondences. Yet Section 3.2.1 and Algorithm 1 (page 14) construct pseudo pairs from mini-batch similarity matrices using Sinkhorn balancing and reciprocal Top-1 matching. The paper does not explain whether target images and texts are sampled independently, how they enter the same candidate pool, or whether original pairing records are used to form the batches. This is an important part of the experimental protocol. If the original correspondence records are used to place true matches in the same batch, candidate matching becomes easier, even if the loss never uses the pairing labels. If the two modalities are sampled independently, the authors need to explain how often true matches are present in a candidate pool.

  2. Figure 3 reports 49,541 pseudo pairs before verification in epoch 1 on R→C, and 64,574 in epoch 3. CUHK-PEDES contains 40,206 images, with two descriptions per image. These numbers do not, by themselves, indicate an error. However, the paper should distinguish unique image-text pairs from matches counted across batches. It should also explain how the one-to-one matching rule handles multiple valid descriptions for the same image. Without this information, it is difficult to assess the quality of target-domain pseudo supervision or compare data-access conditions across methods.

  3. The mathematical form of the DIL bridge loss does not clearly support the proposed intermediate-domain learning mechanism. Equation (10) defines fmix=λsfs+λtftf_{mix} = λ_s f_s + λ_t f_t, with λs+λt=1λ_s + λ_t = 1. Equation (29) in the appendix uses $ L_{bridge} = λ_s ||f_mix − f_s||_2 + λ_t ||f_mix − f_t||_2$.

Substituting Eq. (10) gives $ L_{bridge} = 2 λ_s λ_t ||f_s − f_t||_2 $. For fixed and distinct source and target features, this term is largest when λs=λt=0.5λ_s = λ_t = 0.5. It decreases as either mixing weight approaches zero. Thus, minimizing this loss alone does not favor balanced intermediate-domain mixing. The predictor can instead reduce the loss by moving the mixed feature toward either domain. This does not prove that the weights collapse during full training, since the encoders are also trained with other losses. Still, near-uniform initialization and a bounded Softmax temperature do not rule out this behavior. The appendix also states that the prediction bridge and diversity terms from the original IDM are not used. The authors should explain the gradient flow and show how the mixing weights behave in practice. They should also establish whether the gains come from the proposed intermediate-domain mechanism rather than a general feature-distance constraint.

Questions for authors
  1. How are target images and texts sampled and grouped into batches? Is there a cross-batch feature bank or a global candidate search? Are the original image-text pairing records used at any point when forming batches? Please clarify how the pair counts in Figure 3 are calculated. It would help to report the number of unique images, unique texts, and unique candidate pairs in each epoch. Are target identity labels used only for post hoc evaluation of pseudo-pair accuracy? For CUHK-PEDES, how many valid pairs might be missed because of the one-to-one matching constraint? These details matter for assessing the protocol and the main results.

  2. Could the authors report the distributions of λ_s and λ_t at the start, middle, and end of training? Please also clarify how gradients from Eq. (29) flow through the mixing predictor and the encoders. Two useful controls would be fixed 0.5 mixing weights and removing the bridge loss. If another constraint prevents the mixing weights from becoming extreme, it should be stated in the method section.

  3. Figure 3 reports only the overall pseudo-pair accuracy after filtering and recovery. Could the authors separately report the correct rejection rate and false rejection rate of False-Pair Filtering, as well as the number and identity-level accuracy of pairs added by Missed-Pair Recovery? Hard negatives with similar clothing but different identities are especially relevant. Matching visible attributes does not necessarily mean matching identities. Since the filtering branch checks only selected low-scoring candidates, it would also be helpful to report the error rate among high-similarity pairs.

  4. What exactly is included in the Baseline in Table 4? On R→C, IML raises R@1 from 57.57 to 58.95. Adding DIL raises it to 59.08, and adding TSG raises it to 59.37. Please provide separate ablations for DIL and TSG. In particular, the experiments should distinguish the proposed DIL mixing mechanism from a general feature-distance constraint, and TSG spatial gating from the extra source-domain teacher retrieval loss. These controls would be more informative than adding further transfer directions.

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

Related papers

Discussion (0)

Sign in to comment.