acceptodds
Under review as a conference paper at ICLR 2027

Frequency-Structured Response Consolidation for Lifelong Person Re-Identification

Abstract

Lifelong person re-identification (LReID) learns from new domains while retaining previously acquired knowledge, with catastrophic forgetting remaining its central challenge. Existing data replay methods raise privacy concerns by storing historical exemplars, while knowledge distillation methods face cumulative forgetting of undistilled knowledge. Historical style replay transfers previous domain styles to current images, producing paired current-style and historical-style views and thereby avoiding such privacy concerns. However, since domain styles only partially represent previous domains, only limited historical knowledge can be recovered at each stage. Subsequent domain learning may progressively weaken earlier-domain knowledge, leading to catastrophic forgetting. To address this problem, we propose Frequency-Structured Response Consolidation (FSRC), which combines structured frequency modulation with cross-style response alignment to promote knowledge reuse and consolidation. Specifically, Structured Frequency Modulation (SFM) is embedded into the model to selectively regulate frequency-dependent changes in feature responses induced by historical-style variations, allowing the model to adapt to evolving domain styles and facilitating cross-domain knowledge reuse. Meanwhile, we introduce Cross-Style Alignment (CSA) to align paired feature responses of the same pedestrian under current and historical styles, thereby constraining subsequent representation updates to reduce interference with previously acquired knowledge and support knowledge consolidation. Extensive experiments on the standard LReID benchmark under two training orders demonstrate that FSRC surpasses state-of-the-art methods in both seen and unseen domains.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Person re-identification, Lifelong learning, Frequency modulation, Knowledge consolidation

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

This paper studies exemplar-free lifelong person re-identification. The propose Frequency-Structured Response Consolidation (FSRC), which adds two components at the ResNet-50 backbone. The first component, Structured Frequency Modulation (SFM), transforms the intermediate feature map into the frequency domain, predicts a sample-specific mask that attenuates or amplifies each frequency location, and adds the resulting change back to the feature map through a zero-initialized residual branch. The second component, Cross-Style Alignment (CSA), is used only during training. It encourages the original and historical-style views of the same image to have similar pooled descriptors. The method is evaluated on the standard benchmark with five seen and seven unseen domains under two training orders. FSRC improves over ADSR by roughly one point on the seen-domain average and one to two points on the unseen-domain average.

Critical strengths
  1. The method is well specified, lightweight, and easy to adopt. SFM and CSA are fully described by Eqs. (1)–(9), and all hyperparameters are reported.
  2. The second strength is that the improvements over a strong baseline are consistent in direction, even though they are small. FSRC improves all four final averages over ADSR under both training orders and also improves a locally reproduced DASK, which suggests the idea is not tied to a single replay mechanism.
Critical weaknesses
  1. The reported gains are small, and the paper's own numbers suggest they are within run-to-run noise. All results come from one run with seed 0. No standard deviations or confidence intervals are reported anywhere. Most claimed improvements are below about 1 point.
  2. SFM is closely related to Deep Frequency Filtering[1]. DFF also: (1) applies an FFT to intermediate CNN feature maps. (2) predicts a spatial-attention-style mask in the frequency domain. (3)multiplies the spectrum by this mask.
  3. The claim that regulation is "retained" and "reused" is not substantiated. SFM's parameters are overwritten at each stage like all others, then averaged by model fusion. The paper gives no analysis of how the learned masks evolve across stages, whether masks learned on earlier domains remain meaningful later, or whether SFM's contribution is larger on older domains than on recent ones.
  4. Only a ResNet-50 backbone and a single benchmark protocol are evaluated. Recent LReID work [2,3] increasingly uses ViT or CLIP-based backbones and prompt-based exemplar-free methods. It is unclear whether the method transfers to token-based architectures.

[1] Lin S, Zhang Z, Huang Z, et al. Deep frequency filtering for domain generalization (CVPR 2023)

[2] Xu K, Cheng H, Li J, et al. Vision-Language Attribute Disentanglement and Reinforcement for Lifelong Person Re-Identification (CVPR 2026)

[3] Wen W, Chen H, Zhang S. Prompt-Anchored Vision-Text Distillation for Lifelong Person Re-identification. (CVPR 2026)

Questions for authors

Please report mean ± std over at least 3–5 distinct random seeds under both orders for ADSR, Baseline+SFM, Baseline+CSA, and FSRC. Are the improvements in Tables 1–4 and 8 statistically distinguishable from zero?

Reviewer_tm5sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Frequency-Structured Response Consolidation(FSRC)for exemplar-free lifelong person re-identification. It builds on historical style replay, pairing current images with their historical-style counterparts, and introduces two components. Structured Frequency Modulation(SFM)predicts sample-conditioned spectral masks for intermediate features and applies the resulting corrections through a residual pathway. Cross-Style Alignment(CSA)constrains paired responses through holistic consistency and stability-weighted radial-band objectives. The evaluation uses the LReID benchmark with five seen and seven unseen domains under two training orders. Experiments include comparisons with existing methods, average-forgetting measurements, integration with ADSR and DASK, detailed CSA ablations, and SFM bypass and frequency-position permutation analyses.

Critical strengths
  1. Paired cross-style supervision is meaningfully connected to spectral modulation, with controlled evidence supporting the overall benefit. Using current and historical-style versions of the same image provides a clear correspondence signal for learning style-stable identity representations. In Table 3, the complete method improves the matched baseline’s seen/unseen average mAP from 56.7/66.5 to 58.2/68.3. Table 4 additionally reports improvements when FSRC is incorporated into locally reproduced ADSR and DASK models. These results support the framework’s effectiveness and applicability across different style-replay mechanisms.
  2. Detailed ablations and interventions provide evidence beyond final average accuracy. Table 8 separately examines holistic consistency, radial-band constraints, stability weights, and the separation term. Appendix C.3 further disables SFM within the same checkpoint and permutes the frequency positions associated with modulation values. Together with average-forgetting results under both training orders, these analyses support the continued contribution of learned modulation at inference and provide concrete evidence concerning frequency-position dependence and cross-style consistency.
Critical weaknesses

Critical Weaknesses

  1. Independence between unseen-domain evaluation and hyperparameter selection is not established. Appendix D states that the seven unseen datasets are used only for testing. However, Appendix B.1 motivates the default modulation range, band-alignment weight, and stability temperature directly through unseen-domain performance; Appendix C.1 similarly supports the SFM insertion choice using unseen-domain results. The manuscript does not explain whether these choices were fixed through an independent validation procedure before observing the unseen test results. If test performance informed model selection, the reported improvements would not constitute independent evidence of generalization. This unresolved evaluation issue directly affects a principal conclusion of the paper.
  2. The statistical reliability of the incremental gains and forgetting reductions remains unclear. Section 4.1 states that all experiments use random seed 0, without reporting variation across independent training runs or uncertainty in paired performance differences. Relative to the ADSR results in the main tables, seen-average mAP improves by 1.2 and 0.9 percentage points under the two orders; the Order-1 reductions in AF(mAP) and AF(R@1) are 0.40 and 0.08 points. The two orders provide useful evidence about sequence sensitivity but do not quantify training variability within an order. Consequently, the stability of these increments and some component-level differences remains uncertain, limiting confidence in reliable progress over the strong baseline.
Questions for authors
  1. Clarification—model selection: Were the maximum modulation range, CSA weights, stability temperatures, and SFM insertion position fixed before observing results on the seven unseen test domains? Please specify the validation data and selection criteria, and clarify the unseen-performance-based explanations of the defaults in Appendices B.1 and C.1.
  2. Clarification—result reliability: Beyond the seed-0 results, what independent-run evidence or uncertainty estimates support the stability of FSRC’s improvements over the matched baseline and its reductions in average forgetting? In particular, how large is the variability of the paired performance differences relative to the reported gains?
Reviewer_FY6yOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a style-replay based lifelong person re-identification method based on frequency response modulation. It employs a sample-adaptive mask to selectively attenuate or amplify spectral responses of feature map by historical styles, and align paired original and historical style responses via holistic cross style consistency and stability-weighted band alignment. Experimental results demonstrate that the proposed method outperforms existing ResNet-50-based exemplar-free LReID methods.

Critical strengths

Investigating the modulation of spectral distribution for lifelong person re-id is interesting.

Critical weaknesses
  • Insufficient justification of novel contribution: The novel contribution of the proposed method has not been clearly described compared to existing style replay-based methods and existing frequency-domain learning methods.

  • Insufficient experimental validation: The comparative experimental results are provided with ResNet based baselines only without including more advanced methods, e.g., ViT-based ones and more recent methods.

  • Insufficient discussion of the experimental results: As shown in Tables 1 and 2, the performance of the proposed method is not consistently improved on different training orders compared with the existing methods.

Questions for authors
  • To better understand the novel contribution of the proposed method, it is recommended to explain/discuss existing style replay-based methods in more detail. Moreover, it would be better to clarify the novelty of the proposed SFM compared to the existing frequency-domain learning methods.

  • The comparative experimental results are provided with ResNet based baselines only. It is recommended to compare more advanced methods, e.g., ViT-based ones and more recent methods, such as Long short-term knowledge decomposition and consolidation for lifelong person re-identification (2025 IEEE TPAMI) Distribution-aware Forgetting Compensation for Exemplar-Free Lifelong Person Re-identification (2026 IEEE TMM) AdalReID: Auto-selected Knowledge Adapters for Lifelong Person Re-identification (2026 IEEE TPAMI)

  • As shown in Tables 1 and 2, the performance of the proposed method is not consistently improved on different training orders compared with the existing methods. Discussion is needed.

  • Figure 6 in the supplementary provided parameter analysis on the training order-1. Do we see the same optimal parameters on the training order-2?

  • The proposed FSRC is applicable to existing style replay-based LReID methods as shown in Table 4. Which baseline was used for the proposed method in Tables 1 and 2? Can we still see consistent performance improvement by the proposed method when applied to all the applicable existing baselines in Tables 1 and 2?

  • Regarding SFM, the methodology changes the Fourier transform coefficients instead of changing the frequency itself. Therefore, this reviewer thinks “spectral modulation” would be a more proper terminology compared to “frequency modulation.”

  • DukeMTMC-reID dataset has been officially withdrawn, and hence should not be used for experiments. Experimental results without Duke dataset are need to be provided.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.