acceptodds
Withdrawn from ICLR 2027

Beyond Agreement: Contrastive Evidence Calibration for Composed Image Retrieval

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Composed image retrieval requires preserving relevant reference-image attributes while satisfying a textual modification. Multiple query branches provide complementary retrieval signals, but their agreement may reflect shared appearance cues rather than support for the requested change. We propose Contrastive Evidence Calibration (CEC), a score-level approach that distinguishes branch disagreement, modification matching, reference resemblance, and composed-query compatibility. Starting from a fixed contrastive offset, CEC learns five bounded coefficients over frozen retrieval signals. A regularized full-gallery objective controls the adjustment, while an optional group objective emphasizes distinctions among competing candidates. We characterize the objective in effective coefficient space and derive a sufficient condition for preserving pairwise rankings. With independent training on each dataset, three-seed score ensembles achieve mean recall of 68.82 on the complete FashionIQ validation galleries and 82.43 on the official CIRR test. Matched comparisons show that the fixed offset accounts for the FashionIQ gain, while the complete calibrator improves its CIRR parent by 2.21 points. Applying CIRR’s group objective to automatically constructed FashionIQ groups reduces recall, motivating separate choices for the score representation and its local supervision.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Composed Image Retrieval, Multimodal Retrieval, Vision-Language Models, Contrastive Evidence Calibration, Score Calibration

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_jtwWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper solves the issue of multi-branch retrieval score fusion in Composed Image Retrieval (CIR). The authors argue that score consistency across different retrieval branches does not necessarily imply that a candidate image satisfies the text-based modification requirements, as such consistency might stem merely from shared visual features. To address this, the paper proposes Contrastive Evidence Calibration (CEC), which models score discrepancies across five channels: parent model scores, query branch differences, text-modification matching, reference image similarity, and composed query matching. Building upon fixed score offsets, the method employs five bounded, learnable coefficients for linear adjustment and is optimized using a global gallery retrieval loss, an optional local candidate set loss, and regularization constraints. Experiments were conducted on the FashionIQ and CIRR datasets. CEC achieved an average recall of 68.82 on FashionIQ and 82.43 on the official CIRR test set, representing improvements of approximately 0.31 and 2.21 percentage points, respectively, over the respective parent models. While the results suggest that score adjustment can enhance retrieval performance, the work lacks sufficient innovation, and the reported experiments fail to adequately demonstrate the inherent effectiveness and independent contribution of the proposed learnable calibration mechanism.

Critical strengths
  1. The experimental setup is relatively standardized and provides some valuable controlled comparisons.
  2. Negative results and the limitations of the methods are disclosed quite adequately.
Critical weaknesses
  1. The core method lacks innovation; fundamentally, it remains a form of constrained linear score fusion.
  2. Key ablation studies on the FashionIQ dataset indicate that the learnable calibration module yields no additional performance gains.
  3. CIRR serves as the primary experimental basis for demonstrating the effectiveness of the proposed full calibration method. Table 4 reports that CEC achieves a mean performance improvement of approximately 2.21 percentage points over the parent model. However, this comparison conflates multiple changes—specifically, fixed score offsets, five learnable calibration coefficients, and local candidate-set supervision—without providing sufficient controlled experiments to isolate the individual contribution of each factor.
  4. The method relies on a complex and computationally expensive parent model, raising concerns regarding both practical efficiency advantages and the fairness of experimental comparisons.
  5. The theoretical analysis is rudimentary and fails to establish a substantive link to the core research motivation.
Questions for authors
  1. Table 2 shows that the "fixed offset" approach achieves an Ensemble Mean identical to that of the full CEC, while the learned coefficients in Table 7 are all close to zero. Could the authors explicitly explain what irreplaceable advantage the five learnable coefficients offer over fixed weights on this dataset?
  2. Could the authors provide a complete ablation study for CIRR, comparing the parent model, the fixed offset approach, the version with learnable calibration (using the full gallery only), and the full CEC (incorporating local candidate group supervision)? It would be particularly valuable to understand the individual contribution of each component to the 2.21 percentage point gain and whether these contributions are stable.
  3. Why is there almost no need for learned calibration coefficients on FashionIQ, whereas CIRR requires significant coefficients and local group supervision? Does this suggest that the proposed method fails to establish a stable, universal rule for evidence calibration? Could the authors further analyze the reasons behind this discrepancy?
Reviewer_6eD2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses a failure mode in composed image retrieval where multiple branches favor an incorrect candidate because of shared reference-appearance cues. It proposes Contrastive Evidence Calibration (CEC), which freezes an existing retrieval system and learns five bounded coefficients over branch-score contrasts, modification matching, reference resemblance, and composed-query matching. Training combines full-gallery retrieval, optional candidate-group supervision, and regularization.

Critical strengths
  1. Separately scoring reference resemblance and modification matching is well motivated by the need to preserve relevant attributes while satisfying the requested change.
  2. The CIRR comparison holds representations and ensemble size fixed. The complete method improves the official test Mean from 80.23 to 82.43, with gains in both full-gallery and subset retrieval.
Critical weaknesses
  1. The method's advance over existing score-difference and fusion approaches is limited. SEIZE already combines absolute similarity with differences between pre-edit and post-edit text–candidate matching to emphasize the semantic contribution of the modification [1]. Learned linear combinations of retrieval scores have also been studied extensively [2]. CEC uses different signals and supervision, but Eq. (6) remains constrained linear score fusion. Without comparison against ordinary learned fusion using the same frozen signals, the independent advantages of its contrastive coordinates, initialization, and constraints remain unclear. The two propositions also follow directly from convexity with L2 regularization and the triangle inequality, offering limited additional theoretical insight. 2.The experiments do not establish independent benefits from coefficient learning and group supervision. On FashionIQ, the fixed offset and complete method both achieve 68.82, the learned corrections remain near zero, and the evaluated group objectives do not improve performance. CIRR shows an approximately 2.21% improvement, but only compares the parent with the complete method, without separate results for the fixed offset, global-only calibration, and additional group supervision.
Questions for authors
  1. With identical frozen scores, training objectives, and comparable tuning budgets, how much does CEC improve over regularized linear fusion of the original scores? Which contrastive channels or coefficient constraints are necessary?
  2. Can the authors provide staged CIRR comparisons of the parent, fixed offset, global-only calibration, and complete calibration, reporting both full-gallery and subset metrics?
Reviewer_En9gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Contrastive Evidence Calibration (CEC) for composed image retrieval. It combines five retrieval evidence channels through bounded learnable coefficients and optimizes them using global and optional group-level supervision. Experiments on FashionIQ and CIRR demonstrate improvements over the parent retrieval models, with different effects of calibration and group supervision across datasets.

Critical strengths
  1. The paper investigates the distinction between reference resemblance and modification matching in composed image retrieval.
  2. The proposed calibration introduces minimal additional trainable parameters while retaining the original retrieval representations.
  3. The experiments include matched parent comparisons, multiple training seeds, and detailed ablations on FashionIQ.
Critical weaknesses
  1. The methodological novelty is limited. The proposed method is a linear combination of existing retrieval scores, without introducing a substantially new retrieval mechanism.
  2. The learned calibration provides no additional improvement on FashionIQ. The fixed offset and the complete CEC model achieve almost identical ensemble scores.
  3. The CIRR experiments lack sufficient ablations to isolate the contributions of the fixed offset, learned coefficients, and group supervision. The reported improvement cannot establish which component is responsible.
Questions for authors
  1. Can the authors provide separate CIRR results for the fixed offset, learned calibration without group supervision, and the complete model?
  2. What advantages does the proposed contrastive parameterization offer over standard learned linear score fusion under matched training conditions?

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

Related papers

Discussion (0)

Sign in to comment.