acceptodds
Under review as a conference paper at ICLR 2027

When Samples Are Gone: Keeping Past Knowledge at Work for Lifelong Person Re-Identification

Abstract

Lifelong person re-identification aims to learn from sequential domains while maintaining retrieval performance on previously learned domains after historical samples become unavailable. Current LReID approaches mainly rely on replaying historical samples, transferring prior-model behavior, or reconstructing proxy distributions, three strategies that respectively require access to stored data, depend on current inputs to sufficiently activate old-domain behavior, or preserve historical knowledge only indirectly through proxy observations rather than the learned transformations themselves. To address these limitations, we propose Keeping Past Knowledge at Work (KPKW), an exemplar-free LReID framework that directly retains each domain's learned transformations together with statistical information to guide their subsequent use. KPKW first retains the identity-discriminative transformations learned from each domain as frozen LoRA updates. However, retaining these updates alone does not preserve the historical information needed to guide their later use when jointly applied in a unified encoder. KPKW therefore records layer-wise input second moments and between-class scatter for each frozen update without storing historical samples. The former enables historical response preservation, while the latter provides identity-related directions for subsequent optimization. Guided by these statistics, KPKW learns channel-wise composition coefficients with historical response and output-structure preservation, while soft gradient adjustment further reduces interference with historical identity-related directions. During inference, the learned channel organization coordinates the accumulated frozen updates within a unified encoder, enabling identity representation and retrieval without domain labels. Extensive experiments show that KPKW outperforms existing methods by up to 3.4% on seen domains and 4.8% on unseen domains.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
lifelong learning, person re-identification, exemplar-free learning, knowledge retention, statistical information

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MwC7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Keeping Past Knowledge at Work (KPKW), an exemplar-free framework for lifelong person re-identification. The method preserves domain-specific transformations as frozen LoRA updates and records layer-wise input second moments and between-class scatter before historical data become unavailable. These statistics are subsequently used to coordinate the accumulated LoRA updates through channel-wise composition, historical response and output-structure preservation, and identity-related gradient adjustment.

Critical strengths

S1. A coherent approach to preserving and reusing historical knowledge without exemplar replay.

The paper addresses a meaningful challenge in lifelong ReID: maintaining the utility of previously acquired knowledge when historical images are no longer accessible. Preserving domain-specific transformations together with statistical references is a reasonable design. In particular, the relationship between retained input second moments and historical linear-response discrepancies provides a clear mathematical motivation for the proposed coordination objective.

S2. Competitive empirical performance with relatively extensive supporting analyses.

The method achieves strong results on both seen and unseen domains under two training orders. The supplementary material further includes forgetting analysis, multiple random seeds, component ablations, statistical compression, and computational-cost measurements. These experiments provide useful evidence for the effectiveness of the proposed approach and make its practical trade-offs more transparent.

Critical weaknesses

W1. Insufficient justification for the choice of historical statistics

In Sec. 3.5 (Eqs. 12–15), the authors use the principal directions of between-class scatter (SBS_B) to preserve historical identity-related information. However, SBS_B mainly captures differences between identity means and does not explicitly model within-identity variations, which are important for ReID, as highlighted by DKP (CVPR 2024). The paper does not sufficiently explain why SBS_B is preferred over alternatives such as within-class scatter (SWS_W) or identity-level covariance.

The experimental evidence is also limited. In Table A1 (supplementary, p. 22), the final soft SBS_B projection is compared with alternatives using hard projection, making it difficult to isolate the effect of the chosen statistic. Furthermore, Table A16 (supplementary, p. 36) shows only marginal and inconsistent gains from SBS_B guidance. These results leave the necessity and effectiveness of this particular statistical choice insufficiently established.

W2. Inconsistency in the reported experimental results.

The supplementary material contains an inconsistency between the description on p. 22 and the results in Table A18. The former reports different performance values for the full KPKW configuration from those shown in the table. This raises concerns about the clarity and consistency of the experimental reporting, particularly regarding which runs were used for the efficiency comparisons.

W3. Limited evidence for reduced catastrophic forgetting.

Although KPKW achieves strong final Seen-Avg performance in Tables 2–3 and Fig. 4, higher average accuracy does not necessarily indicate less forgetting. Table A3 reports AF and BWT only for KPKW, without corresponding results from competing methods. As a result, it remains unclear whether the reported improvements mainly come from better historical knowledge retention or stronger initial performance on individual domains.

W4. Limited validation of the historical response constraint.

The proposed response preservation objective is derived from historical layer-input statistics, but its effectiveness in preserving end-to-end identity representations is less clear. Fig. 5 relates the optimized response drift to historical performance, yet this does not directly show whether the original identity features or retrieval relationships are preserved after subsequent learning stages. More direct evidence is needed to support the connection between the proposed statistical constraint and actual historical ReID knowledge retention.

Questions for authors

Q1. Choice of historical statistics (Sec. 3.5, Table A16).

Why do the authors use between-class scatter (SBS_B) rather than alternatives such as within-class scatter (SWS_W) or identity-level covariance, which could better capture within-identity variations? A comparison under the same gradient adjustment strategy and projection strength would help justify this choice. Also, given the limited gains from SBS_B in Table A16, how much does this component actually contribute to preserving historical ReID knowledge?

Q2. Inconsistent results in the supplementary material (p. 22, Table A18).

The results stated for Table A18 on p. 22 of the supplementary material do not match the values reported in the table itself. Could the authors clarify which random seed, checkpoint, and experimental configuration were used for Table A18, and explain the discrepancy?

Q3. Evaluation of catastrophic forgetting (Fig. 4, Tables 2–3, and Table A3).

Fig. 4 and Tables 2–3 mainly report average performance, but a higher Seen-Avg does not necessarily imply less forgetting. Although Table A3 reports AF and BWT for KPKW, corresponding results for competing methods are missing. Could the authors provide a direct comparison under the same protocol to show whether the improvements come from better knowledge retention rather than stronger initial performance on individual domains?

Q4. Validity of the historical response constraint (Fig. 5).

Fig. 5 relates the statistics-weighted response drift to historical-domain performance. However, this drift is also the quantity directly optimized by the proposed method. Could the authors evaluate the actual feature drift or retrieval relationship changes between the original domain-specific models and the final unified model? This would help establish whether the statistical constraint preserves historical identity-discriminative information at the level of the full network.

3 comments
AuthorsOpenReview ↗

Response to Reviewer MwC7 (1/3): Historical statistics and Table A18 correction

Thank you for your questions about the choice of historical statistics and for identifying the Table A18 reporting discrepancy. Our response is provided below:

W1 / Q1. Statistical choice

We use SBS_B to guide coefficient optimization with directions that distinguish historical identity means. During Coordination, the LoRA updates are frozen, and only their channel-wise composition coefficients are optimized. The principal directions of SBS_B describe differences among historical identity means and provide a reference for reducing interference with these directions during composition.

SWS_W describes within-identity variation, which is also important for ReID. However, directions with large within-identity variation do not necessarily distinguish different identities. The choice of statistic therefore depends on the purpose of the adjustment. Here, SBS_B supplies identity-related directions for gradient guidance. Historical knowledge is retained jointly through the frozen LoRA updates, Σ\Sigma-based response and output-structure preservation, and this directional guidance.

We compare the statistics under the same adjustment strategy and strength. To avoid the soft-versus-hard difference in Table A1, we compare no guidance, SBS_B, and SWS_W using the same sequential soft-adjustment rule, β=0.75\beta=0.75, and matched subspace ranks. The remaining losses, initialization, statistical sampling, and training budget are fixed. Each cell reports mAP/R@1.

Guidance sourceOrder-1 SeenOrder-1 UnseenOrder-2 SeenOrder-2 Unseen
None: Response + Output61.9/73.572.9/66.661.4/73.372.9/66.6
SBS_B (ours)62.2/74.273.9/68.061.5/73.372.1/65.4
SWS_W62.1/73.873.6/67.461.2/72.872.8/66.3

Among these configurations, SBS_B is best on all four Order-1 metrics. Under Order-2, it achieves the highest Seen mAP and ties the highest Seen R@1, while no guidance gives the highest Unseen-Avg. Thus, its benefit depends on the training order and evaluation domains.

The corrected Table A16 also clarifies the contribution of SBS_B within the complete method. Checking the experimental logs revealed a copy-and-paste error in the original Full-row accuracy entries. The correct results are:

ConfigurationSeen-Avg mAP/R@1Unseen-Avg mAP/R@1
Response61.9/73.373.3/67.1
Response + SBS_B61.9/73.873.0/66.5
Response + Output61.9/73.572.9/66.6
Full: Response + Output + SBS_B62.2/74.273.9/68.0

Adding SBS_B without Output produces mixed changes. With Output enabled, adding SBS_B improves Seen-Avg by 0.3%/0.7% and Unseen-Avg by 1.0%/1.4%. Thus, on Order-1, SBS_B improves all four retrieval metrics in the complete configuration, and its contribution interacts with output-structure preservation.

W2 / Q2. Table A18 provenance

The Table A18 Discrepancy Is a Copy-and-Paste Error, Not a Seed Difference

The Full-model scores attributed to Table A18 in the overview of Appendix D were incorrectly copied. The correct values are 62.2/74.2 Seen-Avg and 73.9/68.0 Unseen-Avg, in mAP/R@1. Table A18 itself reports these values correctly.

Task-only composition, Response preservation, and KPKW Full in Table A18 all use Training Order-1 and seed 3407. For each configuration, the retrieval accuracy, training time, and Coordination peak memory come from the same training record. The accuracy and resource entries therefore correspond to the same run for each configuration.

AuthorsOpenReview ↗

Response to Reviewer MwC7 (2/3): Acquisition and forgetting

Thank you for highlighting the need to distinguish stronger initial acquisition from better historical knowledge retention. Our response is provided below:

W3 / Q3. Forgetting

We directly compare initial acquisition and subsequent forgetting with PAEMA, DRE, and DKUA under both training orders. To distinguish stronger acquisition from better retention, we report the average performance immediately after learning each of the first four historical domains, their final average performance, AF, and BWT. The fifth domain experiences no subsequent task and is excluded from these four historical-retention measures. Each cell reports mAP/R@1, with AF and BWT expressed as absolute differences (%).

OrderMethodAcquisitionFinal historical performanceAF ↓BWT ↑
Order-1PAEMA68.9/80.861.9/74.47.0/6.4−7.0/−6.4
Order-1DRE74.3/84.561.5/74.012.8/10.6−12.8/−10.6
Order-1DKUA70.5/81.758.0/70.112.5/11.6−12.5/−11.6
Order-1KPKW68.7/81.363.8/78.45.4/3.4−4.9/−3.0
Order-2PAEMA69.3/86.761.4/75.37.9/11.4−7.9/−11.4
Order-2DRE74.7/85.961.1/75.013.6/10.9−13.6/−10.9
Order-2DKUA70.2/81.859.6/72.710.7/9.1−10.7/−9.1
Order-2KPKW70.2/83.463.2/77.97.0/5.5−7.0/−5.5

AF measures the decline from historical peak performance, while BWT measures the change from acquisition-stage to final performance. Both are computed from each method’s stage-by-domain results, rather than using final Seen-Avg as a substitute for forgetting.

KPKW achieves lower AF and higher BWT under both training orders. Its advantage therefore includes better historical knowledge retention and cannot be explained solely by stronger initial acquisition. The frozen updates retain learned transformations, while historical statistical constraints limit changes in their responses during subsequent composition. This comparison provides direct evidence for the retention benefit of this design.

AuthorsOpenReview ↗

Response to Reviewer MwC7 (3/3): End-to-end preservation

Thank you for suggesting direct measurements of end-to-end feature drift and changes in retrieval relationships. Our response is provided below:

W4 / Q4. End-to-end preservation

We directly measure the actual final-feature drift and changes in Query–Gallery retrieval relationships from the original domain-specific models to the final unified model on the same historical images. We quantify these changes using feature cosine drift and Top-10 retrieval overlap, together with historical mAP/R@1 to assess the preservation of identity-discriminative retrieval performance. Each training order is reported separately, with equal-weight averages over its first four domains. Δ\Delta denotes the final unified model’s mAP/R@1 minus that of the corresponding original domain-specific model.

OrderConfigurationFeature cosine drift ↓Top-10 overlap ↑mAP/R@1Δ\Delta mAP/R@1
1No historical statistical constraints0.635248.0%59.6/75.7−13.3/−8.7
1Response preservation0.292755.3%63.8/78.1−9.3/−6.3
1Full KPKW0.300555.0%63.9/78.4−9.2/−6.1
2No historical statistical constraints0.497245.5%57.6/72.9−15.5/−11.9
2Response preservation0.306954.8%62.4/77.0−10.8/−7.8
2Full KPKW0.319954.1%63.2/78.0−10.1/−6.9

Under matched settings, adding response preservation reduces final-feature drift, increases Top-10 retrieval overlap, and improves historical-domain mAP/R@1 compared with composition without historical statistical constraints. Response preservation yields the lowest drift and highest overlap, while Full KPKW achieves the highest mAP and R@1 under both orders. These end-to-end measurements are independent of the optimized response drift in Fig. 5 and support the role of the response constraint in preserving historical identity representations and retrieval relationships.

Reviewer_BhB8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates lifelong person re-identification (LReID), a well-established research problem, and proposes Keeping Past Knowledge at Work (KPKW), an exemplar-free framework built upon a frozen Vision Transformer (ViT) backbone. The proposed method learns domain-specific LoRA updates and leverages stored layer-wise input second-order statistics and between-class scatter information to calibrate channel-wise coefficients, aiming to preserve historical knowledge while adapting to new domains. Experiments under two different five-domain sequences, together with ablation studies, diagnostic visualizations, and sensitivity analyses, are conducted to evaluate the effectiveness of the proposed framework.

Critical strengths

This paper presents a new paradigm for exemplar‑free LReID and addresses a real limitation of previous methods. Instead of replay, knowledge distillation or distribution rehearsal, the method directly keeps frozen domain-specific LoRA update and other aggregate statistics as auxiliary historical references, which is conceptually novel for this task.

This paper states the four stages of the proposed framework clearly: retention, recording, coordination, and retrieval, which facilitates easy following and improves the interpretability of the design.

Critical weaknesses

The proposed framework requires retaining domain-specific LoRA branches and full-matrix statistics, both of which grow linearly with the number of domains. This introduces substantial persistent storage overhead and potentially increases inference FLOPs and latency as more domains are encountered. Although the appendix explores compressed alternatives, these variants incur nontrivial performance degradation. Consequently, the scalability and practical efficiency of the proposed framework remain insufficiently demonstrated.

The experimental evaluation relies partly on baseline results reported in prior publications rather than uniformly reproduced comparisons, raising concerns about experimental fairness and consistency. Furthermore, the evaluation focuses exclusively on exemplar-free methods without including representative sample-replay baselines, limiting the assessment of the method against broader LReID approaches. Additionally, several important supporting experiments are presented only in the appendix, with limited discussion or interpretation in the main manuscript, weakening the empirical justification for the paper's key claims.

Questions for authors

What is the intended stream length and storage budget? As the storage and inference cost of KPKW grows linearly with the number of tasks, what is the maximum number of tasks that the proposed method can reasonably handle in practice? Are there simple strategies to mitigate this linear overhead without hurting performance?

How can the authors guarantee the fairness of comparison without rerunning the baseline results under identical experimental settings? Could the authors provide a protocol-matched rerun of the strongest baselines (at minimum PAEMA, DKUA, and the most recent exemplar-free method) under both training orders?

2 comments
AuthorsOpenReview ↗

Response to Reviewer BhB8 (1/2): Storage, deployment and stream length

Thank you for raising the practical questions of storage growth, inference cost, and longer learning streams. Our response is provided below:

W1 / Q1. Stream length and budget

The five-domain stream follows the sequential LReID evaluation protocol established in prior work. As described in Experimental Setup, we use LPW s2 in place of DukeMTMC-reID. Under this protocol, the full-statistics configuration has a measured final persistent state of 5.707 GiB. Historical statistics are retained for subsequent Coordination but are not loaded during online Retrieval.

KPKW does not impose a preset task-count limit. The five-domain setting is inherited from the evaluation protocol, rather than a capacity constraint of the method. Additional domain updates can be incorporated as learning progresses. Practical capacity is determined by the available storage and computation budgets, since retaining further updates and statistics incurs additional cost. To characterize this scaling, we compare DKUA and Rank-64 KPKW at five and ten stages under Order-1. The ten-stage stream divides each original domain into two identity-disjoint training subsets while preserving the original test sets, seven unseen domains, and total training budget. At each stream length, DKUA’s complete retained state serves as the storage reference. All necessary model and historical states are counted. Matching requires both KPKW’s final state and its maximum stage-wise persistent state to be no greater than DKUA’s. Accuracy is reported as mAP/R@1.

StagesMethodSeen-AvgUnseen-AvgFinal state (GiB)Max stage-wise state (GiB)
5DKUA53.8/64.264.5/57.00.9390.939
5KPKW, Rank-6461.6/73.572.8/66.40.8400.840
10DKUA51.8/62.762.4/54.41.2251.225
10KPKW, Rank-6453.7/68.167.1/59.81.1501.150

A simple function-preserving strategy is to fold the fixed composition into the backbone’s linear layers at deployment. After Coordination, the channel coefficients are fixed, allowing Weffl=W0l+∑iΔWilDiag⁡(mil).W_{\mathrm{eff}}^l=W_0^l+\sum_i\Delta W_i^l\operatorname{Diag}(m_i^l). Precomputing this weight replaces the accumulated branches with a single effective projection per layer. It requires neither retraining nor truncation of the learned transformations and preserves the layer function in exact arithmetic. The resulting online forward therefore uses a fixed number of projections instead of executing every historical LoRA branch individually. We compared the explicit and folded implementations at the same checkpoint and under identical profiling conditions to verify floating-point feature agreement, retrieval accuracy, and runtime. Accuracy is reported as mAP/R@1.

DeploymentSeen-AvgUnseen-AvgMax feature errorLatency (ms/image)Online peak GPU memory (GiB)
Explicit branches62.2/74.073.8/67.3045.6±0.445.6\pm0.40.716
Folded62.2/74.073.8/67.31.6×10−51.6\times10^{-5}6.8±0.16.8\pm0.10.338

The folded model is a separate deployment copy. Independent updates and statistics remain available for subsequent learning, so this conversion addresses online inference and deployment state rather than eliminating continual-learning storage growth. For the latter, Table A6 separately quantifies the accuracy-storage trade-off of statistical compression. We therefore distinguish function-preserving deployment folding from lossy compression of the retained learning state.

AuthorsOpenReview ↗

Response to Reviewer BhB8 (2/2): Baseline fairness, replay and empirical evidence

Thank you for emphasizing fair comparisons, replay baselines, and clear presentation of the empirical evidence. Our response is provided below:

W2 / Q2. Fairness and empirical presentation

Indeed, this is precisely how our comparisons are conducted: KPKW and the baselines are evaluated under the same experimental settings within each of the two evaluation protocols to ensure fairness. As described in Experimental Setup, the comparisons follow Order-1/S6 and Order-2/S7, with matching dataset splits, training sequences, and mAP/R@1 evaluation criteria. Both settings cover five seen domains and seven unseen domains. Tables 2–3 already report the comparisons with PAEMA, DRE, and DKUA under both training orders and distinguish ResNet-based from ViT-based methods.

The evaluation already includes sample-replay methods and is not restricted to exemplar-free approaches. For example, DRE retains historical samples in a memory buffer and uses them for subsequent knowledge preservation. Its results are included in Tables 2–3 under both training orders. The comparison with this ViT-based replay method is summarized below, with each entry reporting mAP/R@1.

MethodO1 Seen-AvgO1 Unseen-AvgO2 Seen-AvgO2 Unseen-Avg
DRE58.7/70.862.9/58.154.6/64.960.0/54.4
KPKW62.2/74.273.9/68.061.5/73.372.1/65.4

KPKW achieves higher seen- and unseen-domain averages under both orders without retaining historical images or instance-level features. Thus, the reported evaluation already provides a direct comparison with a sample-replay baseline.

The supporting experiments were placed in the appendix because of the main-paper page limit. The main manuscript already presents component ablations and historical-response preservation diagnostics. The appendix provides further evidence, including the accuracy–storage trade-off in Table A6, inference costs in Table A9, and KPKW’s stage-wise retention in Table A3. These analyses support the interpretation of the method’s effectiveness and resource requirements. We will consider moving the key results and their accompanying discussion into the main manuscript.

Reviewer_updhOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents KPKW, an exemplar-free framework for lifelong person re-identification. It retains domain-specific LoRA updates on a frozen ViT backbone and records layer-wise input second moments and between-class scatter. A coordination stage then learns channel-wise coefficients over the frozen updates using current-domain supervision and historical statistical constraints. The resulting encoder supports retrieval without domain labels. Experiments under two five-domain training orders report improved average performance on seen and unseen domains over the included baselines.

Critical strengths
  1. Retaining learned updates together with statistical references provides a practical mechanism for coordinating historical knowledge without replaying past samples.
  2. Input second moments make the response discrepancy on recorded layer inputs computable after those inputs are discarded. The paper appropriately distinguishes this property from preservation of end-to-end network behavior.
  3. The evaluation includes seen and unseen domains, component ablations, update-contribution analysis, statistics compression, and training and inference costs.
Critical weaknesses
  1. Positioning relative to prior work could be clearer. The historical response objective is closely related to RegMean [1], while frozen LoRA updates and subspace constraints have precedents in InfLoRA [2]. Targeted comparisons would help isolate the contribution of the proposed coordination scheme.
  2. The additional benefit of the complete configuration remains uncertain. In Table A16, the full model does not outperform response preservation alone. The auxiliary components affect their intended diagnostics, but their incremental retrieval benefits appear modest and variable.
  3. Scalability remains an open consideration. Full statistics require approximately 1 GiB per domain, and the five-domain checkpoint reaches 5.71 GiB. Compression results are encouraging, but longer streams and comparisons under matched resource budgets would better characterize practical scalability.
  4. The appendix distinguishes the selected main run from section-specific controlled runs, but their relationship is not fully clear. In particular, the full-model scores attributed to Table A18 on page 22 differ from those actually reported in Table A18. Please reconcile these entries and identify the run and seed used for each comparison, including whether accuracy and resource measurements come from the same run.
  5. Some related works [3, 4] are not discussed.

[1] InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning. CVPR 2024. [2] Dataless Knowledge Fusion by Merging Weights of Language Models. ICLR 2023. [3] Prototype-guided Knowledge Propagation with Adaptive Learning for Lifelong Person Re-identification. IJCAI 2025. [4] Towards Realistic Lifelong Re-identification: Identity Recurrence with Changing Clothes. ICML 2026.

Questions for authors

Interpretation of channel composition. How should the learned channel-wise coefficients be interpreted? Do they primarily capture shared information across domains, or adjustments specific to the learning sequence?

5 comments
AuthorsOpenReview ↗

Response to Reviewer updh (1/4): Coordination versus RegMean and InfLoRA

Thank you for highlighting the importance of distinguishing KPKW from RegMean and InfLoRA. Our response is provided below:

W1. Relation to RegMean and InfLoRA

(1) RegMean: constrained, supervised composition of frozen updates. RegMean fits merged linear weights by matching source-model responses using input Gram matrices. KPKW shares this response-matching principle, but learns how frozen domain transformations contribute to a unified encoder: Cl=∑i≤tΔWilDiag⁡(mil).C^l=\sum_{i\le t}\Delta W_i^l\operatorname{Diag}(m_i^l). Every ΔWil\Delta W_i^l remains frozen; only the channel coefficients are optimized using current-domain identity supervision together with historical-response and output-structure constraints. The distinction is the constrained parameterization and supervised composition of retained transformations, rather than the second-moment identity itself.

(2) InfLoRA: historical guidance during composition rather than acquisition. InfLoRA uses historical subspaces to design and freeze the input-side factor of a new low-rank branch, while learning its output-side factor. KPKW instead uses historical between-class directions to softly adjust composition-coefficient gradients after the domain updates have been learned and frozen. Thus, its historical guidance acts on the combination of acquired transformations, rather than the learning subspace of a new LoRA branch.

(3) Evidence for coordination. Table A12 compares composition choices within KPKW, in mAP/R@1:

CompositionSeen-AvgUnseen-Avg
Direct sum18.8/29.129.6/22.8
Update-level scalar51.0/66.067.0/59.8
Channel-wise55.3/68.765.4/57.6
Channel-wise + overlap regularization57.8/71.267.8/60.2

These internal controls support learning and structuring the joint use of frozen updates; finer channel granularity alone does not improve every metric.

(4) Matched comparisons. RegMean-style controls share the frozen updates and input statistics; InfLoRA uses a separate ReID adaptation. The response-only pair compares parameterizations under the same objective; the subsequent channel-wise controls examine supervised, structured coordination, while InfLoRA provides a full-strategy comparison. O1/O2 denote the two training orders; each cell is mAP/R@1.

VariantO1 SeenO1 UnseenO2 SeenO2 Unseen
RegMean-style response fitting56.4/67.967.3/61.054.7/66.165.6/58.7
Channel response-only57.3/69.169.9/63.156.7/68.668.3/62.0
Channel task + response + overlap61.8/73.674.0/67.561.5/73.272.0/64.9
KPKW Full62.2/74.273.9/68.061.5/73.372.1/65.4
InfLoRA, ReID adaptation16.1/20.325.8/19.516.8/21.025.6/19.3

Under the matched LReID protocol, KPKW achieves a better overall balance of seen-domain retrieval and unseen-domain generalization than RegMean-style response fitting and ReID-adapted InfLoRA.

AuthorsOpenReview ↗

Response to Reviewer updh (2/4): Not a Lack of Additional Benefit, but a Copy-and-Paste Error in Table A16’s Full Row; channel coefficients

Thank you for your questions about the incremental component benefits and the interpretation of the channel coefficients. Our response is provided below:

W2. Incremental component benefit

Not a Lack of Additional Benefit, but a Copy-and-Paste Error in Table A16’s Full Row

After carefully checking the experimental logs, we confirmed a copy-and-paste error in the Full-model retrieval scores reported in Table A16. The correct values are 62.2/74.2 Seen-Avg and 73.9/68.0 Unseen-Avg. The corrected comparison is shown below, in mAP/R@1:

ConfigurationSeen-AvgUnseen-Avg
Response Preservation61.9/73.373.3/67.1
Response + Output61.9/73.572.9/66.6
Response + SBS_B61.9/73.873.0/66.5
KPKW Full62.2/74.273.9/68.0

Full improves over response preservation alone on all four retrieval metrics, with gains of 0.3%/0.9% on Seen-Avg and 0.6%/0.9% on Unseen-Avg. It also outperforms both Response + Output and Response + SBS_B. These direct improvements in retrieval performance demonstrate the incremental benefit of the complete coordination scheme.

Q1. Coefficient interpretation

The coefficients should be interpreted as channel-wise composition weights learned for a given training sequence, rather than direct measures of information shared across domains. Historical knowledge is retained in the frozen domain-specific LoRA transformations, while the coefficients regulate their use within the unified encoder. They are optimized using current-domain supervision and accumulated historical-statistic constraints to balance current-domain adaptation and historical preservation. Thus, the learning sequence defines the optimization context in which the weights are learned, while cross-domain reuse is the function enabled by the resulting composition.

Table A11 provides functional evidence for this reuse. At the same final checkpoint, we suppress one frozen domain update at a time without retraining or changing the remaining updates and composition coefficients. Removing the Market update reduces Seen-Avg by 4.1%/2.8% and Unseen-Avg by 2.2%/2.2% in mAP/R@1. Removing the CUHK03 update produces decreases of 12.3%/11.7% and 10.5%/10.2%, respectively. Suppressing each of the five updates lowers both Seen-Avg and Unseen-Avg. These results show that the learned composition uses retained transformations for both historical-domain retrieval and unseen-domain generalization.

AuthorsOpenReview ↗

Response to Reviewer updh (3/4): Longer-Stream Scalability and Resource-Matched Comparisons with DKUA

Thank you for suggesting longer-stream and resource-matched evaluations of scalability. Our response is provided below:

W3. Scalability

KPKW maintains higher retrieval performance than DKUA at both five and ten learning stages while using less retained-state storage.

We select DKUA, a strong recent ViT-based baseline included in our main comparisons, for the scalability evaluation. We evaluate the original five-stage stream and a ten-stage stress test constructed by splitting each of the five training domains into two identity-disjoint tasks. This increases the number of learning stages over the same five domains while preserving the original query/gallery sets and seven unseen-domain evaluations. Backbone initialization and total training budgets are matched. KPKW uses rank-64 retained statistics in both settings.

At each stream length, we use DKUA’s measured retained state as the storage-budget reference. The corresponding budgets are 0.939 GiB for five stages and 1.225 GiB for ten stages. KPKW remains within both budgets. Retrieval results are reported as mAP/R@1.

Learning stagesMethodSeen-AvgUnseen-AvgRetained state (GiB)
5DKUA53.8/64.264.5/57.00.939
5KPKW, rank-64 statistics61.6/73.572.8/66.40.840
10DKUA51.8/62.762.4/54.41.225
10KPKW, rank-64 statistics53.7/68.167.1/59.81.150

At ten stages, KPKW improves Seen-Avg by 1.9%/5.4% and Unseen-Avg by 4.7%/5.4%, while using 6.1% less retained-state storage than DKUA. These results show that its retrieval advantage persists with more learning stages and does not depend on retaining more state than the baseline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.