acceptodds
Withdrawn from ICLR 2027

IdioFuse: Identity-Conditioned Score Fusion for Person Re-Identification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Robust person re-identification often combines complementary cues such as face, gait, and body shape. While adaptive fusion typically targets query quality, model strength also varies across identities. We introduce IdioFuse, an identity-conditioned score fusion framework that tailors weights to each gallery identity without training. By contrasting intra-identity consistency against cross-identity impostors, it extracts identity-specific profiles that couple with query-conditioned adaptation via a parameter-free rule. This widens the separation between true and false matches while preserving score calibration. Evaluations on three clothes-changing person re-identification benchmarks show that IdioFuse consistently outperforms statistical, rank-based, and learned baselines, achieving up to an 8.8% absolute reduction in false non-identification rate and demonstrating the value of identity-conditioned fusion in open-set person re-identification.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open-set Person Re-Identification

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

This paper proposes IdioFuse, an identity-conditioned score fusion method for open-set cross-dressing person re-identification. The method uses the similarities of the same identity and across different identities in a labeled gallery to calculate the score intervals and estimate each model’s discriminability for each identity. During inference, IdioFuse first fuses the multi-model scores between the query and individual gallery samples, and then takes the maximum among multiple samples of candidate identities. To support unknown identity rejection, the paper performs shrinkage calibration based on the tail statistics of high-score false recognitions of candidate identities and can combine through normalized product and query-conditioned weights. The authors conduct comparative experiments, ablation studies, and robustness analysis on CCVID, MEVID, and LTCC, with the results generally supporting the effectiveness of identity-conditioned fusion.

Critical strengths
  1. The identity condition weight can utilize the separability information of each person in different recognition models from the gallery, providing intuitive explanation and design motivation. This design does not require additional training of fusion networks and can be combined with existing query condition methods.
  2. The paper covers both video and image benchmarks, and compares multiple fusion methods under the same underlying model. Experiments on identity weight permutation testing, calibration ablation, and gallery expansion provide evidence for the sources of performance and the robustness of the method.
Critical weaknesses

The results for partially open sets fluctuate significantly under different registration identity divisions, and the paper does not further report the differences in paired performance and their uncertainties. Therefore, the stability of performance improvement under these configurations is still unclear. For example, in Table 2, when MEVID is fused with multiple models, the jitter of FNIR for our method is significantly higher than that of other extreme methods.

Questions for authors
  1. The estimated model separability within the gallery may be influenced not only by model-specific features but also by factors such as clothing, camera, and viewpoint. How can we determine whether this statistic reflects identity discriminability that can be transferred to the query side? Is it possible to analyze how identity weights and recognition performance change when the differences in clothing or shooting conditions between the gallery and the query increase?
  2. Table 1 shows that MEVID includes identities with only one gallery sample, in which case the same-identity score statistic cannot be directly calculated according to Eq. (2). What fallback strategy was actually used? Could you provide FNIR under different numbers of gallery samples to illustrate the performance of open-set recognition under few-sample conditions?
  3. Could you provide the paired performance differences and confidence intervals under the same open-set split to assess the differences between metrics and whether the main method's gains are stable?
Reviewer_s5yhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces IdioFuse, which determines identity-adaptive biometric cues to help person retrieval without additional parameters. The proposed method measures intra-/inter-identity similarities to assess each biometric model's faithfulness and assigns the weights for models accordingly. It complements query-conditioned fusion in different way.

Critical strengths

S1. Training-free : The proposed method is training-free, therefore efficient.

S2. Simplicity : The proposed algorithm is clear and intuitive.

Critical weaknesses

W1. Readability : In Introduction, It is somehow difficult to follow for someone who are not familiar to ReID with multiple biometric cues due to the unkind descriptions; e.g., while gallery sets are for testing, are they already labeled? I am not sure that this is a general setting.

W2. Plug-ang-Play compability : While the authors argue that the proposed method is a different line of work compared to query-conditioned fusion (such as QME), the performance improvements upon QME is not clear and even degrades sometimes.

W3. Lack of visualization : The authors should visualize some identity-weight pairs to verify the proposed method works expectedly, but missing.

Questions for authors

Q1. How w^q_k is calculated in Eq 8? the identity of query is unknown, thus can not be computed by Eq 3 since 'genuine' and 'imposter' identities are also unknown.

Q2. Can the authors provide performance comparisons with SOTA clothes-changing ReID methods? while they are fully trained, it may serves as an faithful criterion of maturity.

Q3. Why "Ours(QME)" does not improve the performance clearly?

Q4. Can we see some visualizations of images that represents identity-assigned weights pairs?

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

The paper proposes IdioFuse, an identity-conditioned score fusion framework for open-set person ReID. It estimates per-identity model weights via intra-/inter-identity from the gallery and applies candidate-specific score calibration for open-set rejection, without training a fusion network. Experiments on CCVID, MEVID, and LTCC show consistent gains, especially in open-set FNIR.

Critical strengths
  • The paper presents extensive experiments across three public benchmarks, including various model combinations, baseline comparisons, ablation studies, and sensitivity analyses, demonstrating a substantial experimental effort.
  • The proposed method does not require training an additional fusion network. It derives fusion weights from pretrained models and gallery statistics and can be readily integrated with existing query-adaptive fusion approaches.
Critical weaknesses
  • The expert pool may contain substantial redundancy in identity-discriminative features, so improvements may stem from conventional model ensembling rather than expert complementarity; cross-model correlation analysis and one-backbone-per-modality ablations are missing. Moreover, the paper does not disentangle identity-specific weighting from score calibration; Table 4 shows that calibration alone yields substantial FNIR reductions, so the necessity of identity-conditioned fusion itself remains unclear.
  • The framework requires multiple pretrained models plus gallery-based identity-profile estimation and calibration, but provides no quantitative analysis of FLOPs, inference latency, memory usage, or overhead relative to conventional fusion using the same expert pool. In addition, open-set results show large variance, and no significance tests are reported, making the reliability of the claimed improvements difficult to assess.
Questions for authors
  1. Can the authors further disentangle the contributions of identity-specific weighting and score calibration through controlled ablation experiments?
  2. Beyond the evaluated face, body, and gait recognition models, have the authors explored other types of experts, such as pose-based or attribute-based models, to verify the generalizability of the proposed fusion framework?

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

Related papers

Discussion (0)

Sign in to comment.