acceptodds
Under review as a conference paper at ICLR 2027

Routable Parameter Memories for Exemplar-Free Lifelong Person Re-Identification

Abstract

Lifelong person re-identification (LReID) aims to continuously learn from sequential domains while preserving prior identity-discriminative knowledge. However, substantial inter-domain gaps make the model prone to overwriting earlier representations when adapting to new domains, resulting in catastrophic forgetting. Existing methods typically preserve historical knowledge through image replay, knowledge distillation, or distribution rehearsal, but they respectively face the burden of retaining historical data, insufficient activation of old-domain discriminative cues under substantial domain shifts, and difficulty in directly preserving previously learned identity-discriminative transformations through reconstructed historical visual statistics. To overcome these limitations, we propose Routable Parameter Memories (RPM), an exemplar-free LReID framework built on a parameter-memory perspective. Instead of replaying historical images, previous-model responses, or proxy distributions, RPM directly treats the discriminative parameter updates learned from each domain as parameter memories and draws on biological memory processes in which pattern separation reduces interference among similar or overlapping memory traces and cue-driven pattern completion reinstates relevant memories from partial cues. However, storing parameter memories alone does not determine how new knowledge should be written or how stored knowledge should be read for each input. In this way, the lifelong learning process is organized into the writing, storage, and reading of parameter memories. Specifically, Subspace-Guided Writing uses historical activation subspaces to guide the writing of new memories with reduced overlap, after which the learned domain-specific LoRA updates are frozen and sequentially stored in a parameter memory bank. Finally, during inference, Prototype-Guided Reading uses prototype anchors as retrieval cues to adaptively read and compose relevant memories for each input. Extensive experiments show that RPM outperforms existing methods by up to 5.9% on seen domains and 4.4% on unseen domains, respectively.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Learning, Exemplar-Free Learning, Lifelong Person Re-Identification, Parameter Memory

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_bgrhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Routable Parameter Memories (RPM) for exemplar-free lifelong person re-identification. It introduces two components: Subspace-Guided Writing (SGW) to reduce interference when learning new domain-specific LoRA memories, and Prototype-Guided Reading (PGR) to adaptively retrieve and combine memories for each test image. Experiments show improved retention of previously learned domains and generalization to unseen domains.

Critical strengths

Strength 1: RPM separates the problem into writing, storing, and reading domain-specific LoRA memories. SGW is intended to reduce overlap with historical activation subspaces, while PGR uses prototype similarities to compose memories for each query. This decomposition is easy to understand and directly targets the tension between retaining prior knowledge and adapting to new domains without keeping historical images.

Strength 2: The paper evaluates two domain orders, seen-domain retention, and generalization to unseen datasets, and includes ablations of the writing and reading mechanisms. The reported results show consistent improvements over the listed baselines, while the ablations provide evidence that prototype-guided soft reading and subspace-guided writing contribute complementary benefits.

Critical weaknesses

Weakness 1: SGW shares the idea of using historical activation subspaces to reduce interference with prior tasks, as in GPM [1], while LoDA [2] also explores activation-based subspace decomposition for LoRA continual learning. KeepLoRA [3] is particularly relevant, as it uses residual-subspace projection, SVD-based initialization, and freezes one LoRA factor while training the other. SGW differs in using projected current-domain activation statistics to construct the input directions of new LoRA memories, rather than projecting task gradients to constrain updates. However, the paper does not sufficiently demonstrate the advantages of this design over these related approaches. At the framework level, RPM's incremental parameter memories and input-dependent retrieval are also related to MoRAM [4], although their memory representations and retrieval mechanisms differ. The authors should more clearly establish the distinct contributions of SGW and PGR beyond these existing ideas.

Weakness 2: The evidence on forgetting and efficiency is incomplete. SGW's near-zero subspace overlap does not necessarily imply less forgetting on previously learned domains. Moreover, some AF/BWT results in the appendix are reconstructed estimates rather than direct measurements, making their reliability difficult to assess without the underlying per-domain, per-stage results.

Weakness 3: RPM also has substantially higher inference latency than some baselines, raising concerns about its computational efficiency as the memory bank grows.

Questions for authors

Question 1: How does SGW differ technically from GPM [1], LoDA [2], and KeepLoRA [3] in constructing and constraining the LoRA writing subspace? In particular, what advantages does SGW's activation-based subspace projection and SVD-based LoRA-A initialization over existing subspace-constrained LoRA methods? Can the authors provide direct or appropriately adapted comparisons under the same ReID evaluation protocol?

Question 2: How does RPM's domain-level memory design and prototype-guided retrieval differ from MoRAM's [4] rank-1 memory expansion and self-activated retrieval, and what advantages does it provide?

Question 3: Forgetting evaluation and efficiency. Can the authors provide actual per-domain, per-stage performance matrices underlying the AF/BWT metrics?

Question 4:How does inference latency scale with the number of stored memories, and can sparse Top-k retrieval improve efficiency without substantially compromising accuracy?

References

[1] G. Saha, I. Garg, and K. Roy. "Gradient Projection Memory for Continual Learning." ICLR, 2021. https://openreview.net/forum?id=3AOj0RCNC2

[2] L. He, D. Cheng, H. Wang, X. Yang, N. Wang, and X. Gao. "Task-Driven Subspace Decomposition for Knowledge Sharing and Isolation in LoRA-based Continual Learning." ICML, 2026. https://proceedings.mlr.press/v306/he26f.html

[3] "KeepLoRA: Continual Learning with Residual Gradient Adaptation." ICLR, 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/dc3ca8bcd613e43ce540352b58d55d6d-Abstract-Conference.html

[4] H. Lu, C. Zhao, J. Xue, L. Yao, K. Moore, and D. Gong. "Little By Little: Continual Learning via Incremental Mixture of Rank-1 Associative Memory Experts." ICML, 2026. https://proceedings.mlr.press/v306/lu26au.html

9 comments
AuthorsOpenReview ↗

Response to Reviewer bgrh (1/4): SGW writing subspaces compared with GPM, LoDA, and KeepLoRA

Thank you for asking how SGW differs from GPM, LoDA, and KeepLoRA in constructing and constraining writing subspaces. Our response is provided below:

W1 / Q1. Construction, constraints, and advantages of the writing subspace.

SGW constructs current-domain-informed input directions before optimization and maintains the selected subspace by freezing LoRA-A. With all LoRA updates disabled, we compute the current-domain activation statistic CtlC_t^l, project it using the historical writing state as C~tl=Pt−1lCtl\widetilde C_t^l=P_{t-1}^lC_t^l, and initialize LoRA-A from its leading left singular directions. LoRA-A remains fixed while LoRA-B is optimized.

The technical distinctions from the three related approaches are:

  • GPM derives protected bases from historical activations and projects gradients during optimization. SGW uses the projected current-domain statistic to specify a low-rank input subspace at initialization.
  • LoDA constructs general and isolated subspaces by optimizing total old/new projection energy and new-to-old relative projection energy, respectively, and fixes the down-projections of its dual branches. SGW selects writing directions through historical-subspace projection of the current statistic.
  • KeepLoRA performs SVD on the projected initial task gradient, protecting both pre-trained weight principal directions and historical feature directions. SGW instead uses projected current-domain activation statistics and an activation-based historical writing state.

The distinction lies in how the writing directions are selected and constrained. SGW combines current-domain statistical information with historical directional constraints; fixing LoRA-A prevents the selected input subspace from drifting during optimization. Table 3 supports its contribution within RPM: adding SGW to the PGR configuration improves Seen-Avg from 64.1/76.0 to 65.2/76.7 and Unseen-Avg from 72.7/65.9 to 73.3/67.6, in mAP/R@1.

We have added ReID-adapted comparisons with GPM, LoDA, and KeepLoRA under the same evaluation protocol. The comparison uses the same ImageNet-1K-pretrained TransReID ViT-B/16 initialization, two training orders, five seen domains, seven unseen domains, query/gallery splits, identity-classification and triplet objectives, epoch budget, and ten paired training seeds. Hyperparameters are selected through independent source-domain validation. Each adaptation preserves the corresponding method’s defining constraints, optimization, and update-integration rules, with actual trainable parameters and retained states reported.

These method-level ReID comparisons directly evaluate the practical performance of the different approaches under a common task protocol. Together with the internal ablation in Table 3, they provide complementary evidence for assessing SGW’s contribution. Each entry below reports mean ± standard deviation over ten training seeds, in mAP/R@1 format:

MethodOrder-1 Seen-AvgOrder-1 Unseen-AvgOrder-2 Seen-AvgOrder-2 Unseen-Avg
GPM, ReID-adapted53.6±0.7/66.1±0.553.6\pm0.7 / 66.1\pm0.561.0±0.9/53.9±0.861.0\pm0.9 / 53.9\pm0.854.9±0.6/67.0±0.454.9\pm0.6 / 67.0\pm0.461.9±0.8/55.3±0.761.9\pm0.8 / 55.3\pm0.7
LoDA, ReID-adapted52.9±0.8/66.7±0.652.9\pm0.8 / 66.7\pm0.668.0±0.7/61.2±0.968.0\pm0.7 / 61.2\pm0.953.3±0.9/66.0±0.553.3\pm0.9 / 66.0\pm0.566.0±1.0/58.6±0.866.0\pm1.0 / 58.6\pm0.8
KeepLoRA, ReID-adapted55.3±0.5/67.0±0.455.3\pm0.5 / 67.0\pm0.466.8±0.6/60.2±0.766.8\pm0.6 / 60.2\pm0.754.5±0.7/67.4±0.554.5\pm0.7 / 67.4\pm0.566.2±0.8/59.7±0.666.2\pm0.8 / 59.7\pm0.6
Ours65.20±0.11/76.73±0.2665.20\pm0.11 / 76.73\pm0.2673.31±0.37/67.62±0.5873.31\pm0.37 / 67.62\pm0.5864.50±0.10/75.66±0.2164.50\pm0.10 / 75.66\pm0.2172.60±0.35/66.38±0.4372.60\pm0.35 / 66.38\pm0.43
AuthorsOpenReview ↗

Response to Reviewer bgrh (2/4): Domain-level memories and prototype-guided reading compared with MoRAM

Thank you for asking about the differences between RPM and MoRAM in memory organization and retrieval. Our response is provided below:

W1 / Q2. Domain-level memory and prototype-guided retrieval versus MoRAM.

RPM and MoRAM differ in both memory granularity and the source of their retrieval weights. MoRAM expands rank-1 key–value atoms and obtains token-level mixture weights through sparse self-activation of their intrinsic keys. RPM preserves each domain’s learned, multilayer LoRA transformations as a domain-level memory. Multiple prototype anchors, computed with LoRA disabled, provide visual references for addressing that memory.

For each image, RPM computes a query in the same frozen-backbone feature space as the anchors. Prototype similarities produce image-level soft weights shared across the inserted layers, while layer-wise amplitudes modulate each memory’s contribution at different depths. This separates visual addressing from the LoRA factors used for constrained writing. The input factors can serve the writing constraint, while the anchors provide independent reading cues. Each domain’s learned transformation remains available as a unit for image-dependent composition.

Table 5 supports the proposed reading choice within RPM: multi-anchor soft reading achieves 65.2/76.7 Seen-Avg and 73.3/67.6 Unseen-Avg, compared with 61.0/72.3 and 52.8/45.2 for hard prototype reading. Single-anchor reading reaches 64.7/76.1 and 72.4/66.1, supporting the use of multiple visual references per domain.

We have additionally compared our method with ReID-adapted MoRAM under the same protocol, preserving MoRAM’s rank-1 memory expansion and sparse intrinsic-key self-activation. Both methods produce retrieval embeddings without ground-truth test-domain labels. The Ours row uses the same matched evaluation as in Q1.

The comparison uses the same ImageNet-1K-pretrained TransReID ViT-B/16 initialization, two training orders, five seen domains, seven unseen domains, query/gallery splits, identity-classification and triplet objectives, epoch budget, and ten paired training seeds. Hyperparameters are selected through independent source-domain validation.

This comparison directly evaluates the practical effects of domain-level memories with separate prototype addressing versus rank-1 memories with intrinsic-key self-activation, complementing the reading-strategy ablations in Table 5. Each entry reports mean ± standard deviation over ten training seeds, in mAP/R@1 format:

MethodOrder-1 Seen-AvgOrder-1 Unseen-AvgOrder-2 Seen-AvgOrder-2 Unseen-Avg
MoRAM, ReID-adapted30.4±0.8/43.2±0.630.4\pm0.8 / 43.2\pm0.644.6±0.9/38.7±0.744.6\pm0.9 / 38.7\pm0.725.8±1.0/38.8±0.825.8\pm1.0 / 38.8\pm0.839.4±0.7/32.6±0.939.4\pm0.7 / 32.6\pm0.9
Ours65.20±0.11/76.73±0.2665.20\pm0.11 / 76.73\pm0.2673.31±0.37/67.62±0.5873.31\pm0.37 / 67.62\pm0.5864.50±0.10/75.66±0.2164.50\pm0.10 / 75.66\pm0.2172.60±0.35/66.38±0.4372.60\pm0.35 / 66.38\pm0.43
AuthorsOpenReview ↗

Response to Reviewer bgrh (3/4): Actual stage-by-domain performance matrices for AF/BWT

Thank you for requesting the stage-by-domain performance matrices underlying AF/BWT. Our response is provided below:

W2 / Q3. Actual per-domain, per-stage performance matrices.

We have conducted additional stage-by-domain retrieval evaluations for RPM under both training orders. Below, we provide the actual performance matrices used to calculate AF/BWT. Each row corresponds to the model after completing a training stage, and each column corresponds to a domain. Each entry reports that stage model’s mAP/R@1 on the corresponding domain. “n/a” denotes a domain not yet learned.

RPM, Training Order-1: Market → SYSU → LPW-s2 → MSMT17-V2 → CUHK03.

StageMarketSYSULPW-s2MSMT17-V2CUHK03
182.0/92.8n/an/an/an/a
274.9/89.290.0/90.6n/an/an/a
371.8/89.295.0/92.670.0/86.0n/an/a
469.6/87.589.0/90.075.0/80.052.3/84.4n/a
569.7/86.989.7/90.861.0/71.945.5/71.859.9/62.0

RPM, Training Order-2: LPW-s2 → MSMT17-V2 → Market → SYSU → CUHK03.

StageLPW-s2MSMT17-V2MarketSYSUCUHK03
149.0/65.0n/an/an/an/a
256.5/70.029.3/61.0n/an/an/a
358.8/73.039.9/69.963.0/84.2n/an/a
455.8/72.537.0/67.078.2/93.391.3/94.7n/a
556.8/69.038.4/65.977.2/90.390.6/91.759.5/61.6

AF/BWT are calculated directly from these measured results. RPM achieves low AF and favorable BWT under both training orders while maintaining strong final retrieval performance. These results demonstrate RPM’s effectiveness in mitigating forgetting and preserving previously learned knowledge during sequential learning.

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

This paper proposed an exemplar-free lifelong person re-identification method, that learns domain-specific LoRA parameters as Routable Parameter Memories (RPM) inspired by biological neuron’s behavior. It uses historical activation subspaces to guide the memory writing, and composes stored parameters guided by prototypes for lifelong learning. Experimental results tested on two training orders of five benchmark datasets show that the proposed method outperforms existing methods in terms of mAP and R@1 scores.

Critical strengths
  • The idea of using domain-specific low-rank parameters for lifelong learning is interesting.

  • Memory writing and reading methods are reasonably designed.

Critical weaknesses
  • Insufficient comparison in terms of memory usage: Though the authors claim the proposed method is exemplar-free, this methodology requires incrementally increasing memories to store LoRA parameters as well as prototype anchors, as new domains come. Therefore, a fair comparison between the two methodologies should be needed considering the additional memory required to store exemplar images and the additional memory required to store parameters and prototypes.

  • Insufficient comparison in terms of computational complexity: In the proposed method, the current activation statistics are estimated by current-domain samples using all token activations at each inserted layer, which may substantially increase the computational complexity in memory writing. Moreover, the proposed method increases the computational complexity of memory reading due to the comparison of the current image with stored prototype anchors. It is recommended to provide comparative experimental results of the actual computational complexity.

  • Incomplete ablation study: The ablation study should be performed considering the memory usage and the computational complexity together, since omitting some modules avoids additional memory uage and computational complexity.

  • Insufficient comparison with more recent methods such as Kunlun Xu et al., “Long short-term knowledge decomposition and consolidation for lifelong person re-identification,” IEEE TPAMI, 2025. Shiben Liu et al., “Distribution-aware forgetting compensation for exemplar-free lifelong person re-identification,” IEEE TMM, 2026

  • Insufficient discussion: The discussion in the experimental section is too simple to understand the behavior of the proposed method sufficiently, e.g., Section 4.2. It would be better to discuss the reason why the performance of the proposed method is not consistently improved compared with the existing methods in Tables 1 and 2. Moreover, it is recommended to discuss the reason why the proposed method provides relatively worse performance at early domains compared to some existing methods. If space is limited, Figure 4 can be removed, as it merely visualizes the data from Tables 1 and 2 in a different format without providing any additional insights.

Questions for authors
  • Insufficient comparison in terms of memory usage: Though the authors claim the proposed method is exemplar-free, this methodology requires incrementally increasing memories to store LoRA parameters as well as prototype anchors, as new domains come. Therefore, a fair comparison between the two methodologies should be needed considering the additional memory required to store exemplar images and the additional memory required to store parameters and prototypes.

  • Insufficient comparison in terms of computational complexity: In the proposed method, the current activation statistics are estimated by current-domain samples using all token activations at each inserted layer, which may substantially increase the computational complexity in memory writing. Moreover, the proposed method increases the computational complexity of memory reading due to the comparison of the current image with stored prototype anchors. It is recommended to provide comparative experimental results of the actual computational complexity.

  • Incomplete ablation study: The ablation study should be performed considering the memory usage and the computational complexity together, since omitting some modules avoids additional memory uage and computational complexity.

  • Insufficient comparison with more recent methods such as Kunlun Xu et al., “Long short-term knowledge decomposition and consolidation for lifelong person re-identification,” IEEE TPAMI, 2025. Shiben Liu et al., “Distribution-aware forgetting compensation for exemplar-free lifelong person re-identification,” IEEE TMM, 2026

  • Insufficient discussion: The discussion in the experimental section is too simple to understand the behavior of the proposed method sufficiently, e.g., Section 4.2. It would be better to discuss the reason why the performance of the proposed method is not consistently improved compared with the existing methods in Tables 1 and 2. Moreover, it is recommended to discuss the reason why the proposed method provides relatively worse performance at early domains compared to some existing methods. If space is limited, Figure 4 can be removed, as it merely visualizes the data from Tables 1 and 2 in a different format without providing any additional insights.

4 comments
AuthorsOpenReview ↗

Response to Reviewer NCfb (1/4): Additional storage and retrieval performance compared with DRE

Thank you for suggesting a comparison of additional storage and retrieval performance with exemplar-replay methods. Our response is provided below:

W1 / Q1. Storage fairness with exemplar replay.

We directly compare our method with exemplar-replay methods in terms of additional storage and retrieval performance, using DRE (TNNLS 2025), a recent and strong exemplar-replay method, as the comparator. We apply the same accounting boundary to both methods, counting retained images, extra model/adapter parameters, prototypes, and auxiliary historical states. One active base model is reported separately; any additional model required across stages is counted as historical storage.

Our method’s five-domain memory bank occupies 182.111 MiB, comprising 181.875 MiB of LoRA parameters, 0.235 MiB of prototype anchors, and approximately 0.001 MiB of layer-wise amplitudes (Appendix E.5). The storage is dominated by the LoRA parameters, with only a small overhead from anchors and amplitudes. Under the reported configuration, each additional domain contributes approximately 36.422 MiB to the bank. The shared writing state (Appendix C.8) is counted separately and included in the additional-storage total below.

After five domains, DRE retains 2,500/2,500 replay images in Order-1/Order-2. We use a consistent storage-accounting convention for both methods: images are counted as uncompressed RGB, and numerical states are counted at FP32. Backbone and pretraining settings are specified for each configuration. The table reports additional historical storage in MiB and retrieval performance as mAP/R@1. Each row pairs storage with performance from the same configuration.

OrderMethodHistorical images (MiB)Extra parameters/reading states (MiB)Auxiliary historical states (MiB)Additional storage (MiB)Seen-AvgUnseen-Avg
1DRE (2025)937.50012.8210.364950.68558.7/70.862.9/58.1
1Ours0189.4358.744198.17965.2/76.773.3/67.6
2DRE (2025)937.50012.8210.364950.68554.6/64.960.0/54.4
2Ours0189.4358.893198.32864.5/75.773.5/67.8

Compared with exemplar-replay methods, our method reduces additional historical storage by approximately 79% while achieving higher Seen-Avg and Unseen-Avg mAP/R@1 in both training orders.

AuthorsOpenReview ↗

Response to Reviewer NCfb (2/4): Complete reading costs; stage-level writing overhead and SGW comparison

Thank you for asking about the actual computational costs of memory writing and reading. Our response is provided below:

W2 / Q2. Actual computational cost of memory writing and reading.

The complete two-pass reading pipeline and prototype-based weighting are included in our inference-cost comparison. Appendix E.5 compares complete inference under the same hardware and evaluation protocol. On Training Order-1, our method requires 57.8 G FLOPs/image and 64.66±0.4164.66\pm0.41 ms/image, using FP32, batch size 1, 100 warm-up iterations, and 1,000 measured images. Latency is reported as mean ± standard deviation over three runs.

Prototype-guided reading accounts for 10.5% of the profiled inference time. The five-memory component profile reports:

ComponentLatency (ms/image)Fraction (%)
LoRA-free query extraction6.77810.6
Prototype-guided reading6.75010.5
Routed feature extraction50.66078.9

Routed feature extraction accounts for most of the profiled time. Both the query pass and prototype-based weighting are explicitly measured and included in the complete reading cost.

For writing, we separately quantify activation-statistics collection, subspace processing, and anchor construction. Statistics collection uses a 2,000-image sampling budget, with all token vectors of the sampled images contributing at each hooked layer (Appendix C.1). Statistics collection and subspace processing occur at stage initialization and historical-state update, rather than at every optimization step. Anchor feature extraction and clustering are also recorded separately.

Writing operations are timed using synchronized wall-clock measurements covering CPU and GPU operations. Compared configurations use the same hardware, precision, and timing scope. The following times are five-stage totals for each order; statistics collection includes initialization and historical-state update, and shared operations are counted once. Total learning time is measured directly over the complete learning procedure, including all required preparation, optimization, and state updates, with evaluation reported separately.

OrderStatistics (s)SVD/state (s)Anchors (s)Optimization (s)Total learning (s)
1421.6717.341.728,307.929,640.1
2367.6602.553.534,002.435,298.1

Statistics collection, SVD/state processing, and anchor construction together account for 3.98% and 2.90% of total learning time in Order-1 and Order-2.

To directly quantify the additional cost of Subspace-Guided Writing, we compare the full method with a version that removes SGW while retaining prototype-guided reading on Order-1. Both configurations use the same backbone, LoRA rank, training sequence, and training budget. Without SGW, total learning time is 24,686.6 s; enabling SGW increases it by 20.1% (29,640.1 s). Anchor construction is retained and included in the cost of both configurations.

Table 3 of the main paper directly compares “+ Prototype-Guided Reading” (without SGW) with “RPM (Full)” (with SGW) on Training Order-1. Adding SGW improves Seen-Avg by 1.1 mAP and 0.7 R@1 percentage points, and Unseen-Avg by 0.6 mAP and 1.7 R@1 percentage points. Combined with the measured 20.1% increase in total learning time for these two configurations, these results demonstrate the retrieval gains and associated learning-time cost of SGW. Auxiliary writing operations account for 3.98%/2.90% of total learning time in Order-1/Order-2, while prototype-guided reading accounts for 10.5% of the profiled inference time. These measurements support a favorable performance–computation trade-off: auxiliary writing overhead is limited, and prototype-guided reading is not the dominant inference bottleneck.

AuthorsOpenReview ↗

Response to Reviewer NCfb (3/4): Resource-aware component ablation

Thank you for suggesting that the component ablation also account for storage and computation. Our response is provided below:

W3 / Q3. Component-wise storage and computational costs.

We compare the five configurations in Table 3 on Training Order-1, pairing their retrieval accuracy with additional retained storage, total learning time, and complete inference FLOPs/latency. Additional storage excludes the common backbone and includes all states retained for inference or future learning. Total learning time covers all five stages, including preparation, optimization, and state updates; evaluation is timed separately.

All configurations use the same backbone, LoRA rank, training budget, hardware, and precision. Online inference uses FP32, batch size 1, 100 warm-up iterations, and 1,000 measured images, with latency reported as mean ± standard deviation over three runs. Each variant’s complete feature-extraction path is measured, including the additional query forward where applicable. Accuracy values below are from Table 3.

ConfigurationSeen mAP/R@1Unseen mAP/R@1Additional storage (MiB)Total learning (s)Inference FLOPs (G) / latency (ms)
Continual LoRA58.4/72.169.8/63.243.69920,691.425.2 / 14.63±1.7714.63\pm1.77
Fixed All-Memory58.8/72.270.7/64.0189.19922,536.635.1 / 37.12±1.2937.12\pm1.29
+ SGW60.8/73.570.4/63.5198.05723,036.135.1 / 37.32±0.4937.32\pm0.49
+ PGR64.1/76.072.7/65.9189.43624,686.657.8 / 60.07±0.9360.07\pm0.93
Our full method65.2/76.773.3/67.6198.17929,640.157.8 / 60.26±0.6760.26\pm0.67

Our full method achieves the best Seen-Avg and Unseen-Avg retrieval performance among all five configurations. SGW and PGR provide complementary retrieval gains with limited additional storage and computation, yielding a favorable performance–resource trade-off.

AuthorsOpenReview ↗

Response to Reviewer NCfb (4/4): Recent-method comparisons; average gains and domain-wise trade-offs; shared background wording

Thank you for your comments on recent methods, domain-wise performance, and shared introductory wording. Our response is provided below:

W4 / Q4. Comparison with recent methods.

The cited TPAMI 2025 method is already included as LSTKC++ in Tables 1–2. Its results are reproduced below alongside our method and DAFC, the cited TMM 2026 method.

The comparison with DAFC is summarized below using the training orders and evaluation format of the main tables. Each entry reports mAP/R@1 for the five seen domains, Seen-Avg, and Unseen-Avg. DAFC uses a CLIP-pretrained ViT-B/16, while our method uses an ImageNet-1K-pretrained TransReID ViT-B/16; the backbone types are explicitly listed.

Training Order-1: Market → SYSU → LPW → MSMT17 → CUHK03.

MethodBackboneMarketSYSULPWMSMT17CUHK03Seen-AvgUnseen-Avg
LSTKC++ (TPAMI 2025)ResNet67.5/83.986.3/88.252.1/61.919.4/40.347.0/47.854.5/64.462.9/55.6
DAFC (TMM 2026)CLIP ViT-B/1672.8/87.387.6/89.150.6/59.615.2/33.050.4/52.155.3/64.265.1/58.5
OursViT-B/1669.7/86.989.7/90.861.0/71.945.5/71.859.9/62.065.2/76.773.3/67.6

Training Order-2: LPW → MSMT17 → Market → SYSU → CUHK03.

MethodBackboneLPWMSMT17MarketSYSUCUHK03Seen-AvgUnseen-Avg
LSTKC++ (TPAMI 2025)ResNet55.7/65.715.0/34.366.9/83.686.3/87.648.4/49.954.4/64.263.4/56.0
DAFC (TMM 2026)CLIP ViT-B/1660.8/70.519.3/39.762.7/82.587.5/89.153.6/55.456.8/67.466.1/59.3
OursViT-B/1656.8/69.038.4/65.977.2/90.390.6/91.759.5/61.664.5/75.773.5/67.8

Our method achieves higher Seen-Avg and Unseen-Avg mAP/R@1 than both compared methods under both training orders, demonstrating stronger seen-domain retention and unseen-domain generalization.

W5 / Q5. Overall performance and domain-wise trade-offs.

Our primary evaluation criteria are average retention across the five seen domains and average generalization across the seven unseen domains. Domain-wise rankings vary across the compared methods, and no method dominates every domain. For example, PAEMA achieves higher Market mAP than our method in Order-1 (73.1 vs. 69.7), while our method performs better on MSMT17 (45.5 vs. 30.4). Similarly, DKUA performs better on LPW in Order-2 (66.2 vs. 56.8), whereas our method performs better on Market (77.2 vs. 61.7). Our method nevertheless achieves higher Seen-Avg and Unseen-Avg under both training orders in Tables 1–2. The early-domain gaps are therefore part of the domain-wise performance trade-offs; our main conclusion concerns stronger overall retention and generalization, with these individual-domain differences explicitly reported.

The relevant averages from Tables 1–2 are summarized below, with each entry reporting mAP/R@1:

MethodOrder-1 Seen-AvgOrder-1 Unseen-AvgOrder-2 Seen-AvgOrder-2 Unseen-Avg
PAEMA60.3/70.269.8/63.260.9/70.570.5/63.8
DKUA59.5/69.567.8/60.360.4/71.366.9/59.7
Ours65.2/76.773.3/67.664.5/75.773.5/67.8

The existing reading ablation also supports the overall benefit of multi-memory soft composition. In Table 5, Soft Prototype Reading achieves 65.2/76.7 on Seen-Avg and 73.3/67.6 on Unseen-Avg, compared with 61.0/72.3 and 52.8/45.2 for Hard Prototype Reading. This supports cross-domain memory composition as a useful design choice for overall retention and generalization, although individual datasets can favor different methods.

Figure 4 does not provide independent stage-wise evaluation evidence. Its curves were constructed using final-table results, rather than independent evaluations after each training stage. We agree that the discussion should prioritize the domain-wise comparisons and ablation evidence above.

E1. Shared background wording.

The quoted overlap is in introductory background wording; it does not imply that the two papers are substantively the same. The passage defines cross-camera ReID and introduces the motivation of evolving deployment environments, rather than describing the proposed method or its technical contributions.

RPM’s contribution lies in the combination of subspace-guided memory writing, frozen domain-specific LoRA memories, and prototype-guided input-dependent reading. Similarity in this general background description does not establish that the two papers share the same method, experiments, or substantive contributions.

Comments on the paper

6 comments
AuthorsOpenReview ↗

Response to Reviewer aA2h (1/3): Measured AF/BWT and per-target results; rationale for rank 64 and selection scope

Thank you for raising the questions about evaluation provenance, forgetting statistics, per-target results, and configuration selection. Our response is provided below:

W1. Measured results and configuration selection.

(1) Evaluation provenance and measured results. Tables A8–A11 are based on actual evaluations. Specifically, the AF/BWT values in Tables A8/A9 are calculated from recorded per-domain, per-stage performance, rather than inferred from cumulative Seen-Avg trajectories. Tables A10/A11 contain direct per-target evaluations of RPM and the compared baselines. The original captions inaccurately described the provenance of these results. We have corrected the wording in Appendices E.7–E.8; the numerical results are unchanged.

AF/BWT. RPM’s final Seen-Avg and forgetting statistics are:

Training orderSeen-Avg mAP/R@1AF mAP/R@1 (pp)BWT mAP/R@1 (pp)
Order-165.2/76.79.6/8.6-7.1/-8.1
Order-264.5/75.71.3/3.57.6/3.0

RPM achieves the highest final Seen-Avg under both training orders. In Order-1, its AF of 9.6/8.6 is lower than PAEMA’s 14.8/14.3, DRE’s 10.2/9.8, and DKUA’s 12.1/11.4. In Order-2, RPM achieves AF of 1.3/3.5 and positive BWT of 7.6/3.0, together with the highest final Seen-Avg of 64.5/75.7. These measured results support RPM’s ability to preserve previously learned knowledge while maintaining strong final retrieval performance.

Measured performance on individual unseen targets. The following results reproduce the direct evaluations reported in Tables A10/A11. Each entry is mAP/R@1.

Training Order-1

MethodCUHK01CUHK02VIPeRPRIDi-LIDSGRIDSenseReIDUnseen-Avg
PAEMA82.0/81.679.9/79.164.8/54.572.9/64.683.4/77.356.1/46.049.5/39.369.8/63.2
DRE75.1/76.573.0/74.057.9/49.466.0/59.576.5/72.249.2/40.942.6/34.262.9/58.1
DKUA80.0/78.777.9/76.262.8/51.670.9/61.781.4/74.454.1/43.147.5/36.467.8/60.3
RPM85.5/86.083.4/83.568.3/58.976.4/69.086.9/81.759.6/50.453.0/43.773.3/67.6

Training Order-2

MethodCUHK01CUHK02VIPeRPRIDi-LIDSGRIDSenseReIDUnseen-Avg
PAEMA82.5/82.681.7/80.964.5/53.373.7/66.084.6/77.755.7/44.850.8/41.370.5/63.8
DRE72.0/73.271.2/71.554.0/43.963.2/56.674.1/68.345.2/35.440.3/31.960.0/54.4
DKUA78.9/78.578.1/76.860.9/49.270.1/61.981.0/73.652.1/40.747.2/37.266.9/59.7
RPM85.5/86.684.7/84.967.5/57.376.7/70.087.6/81.758.7/48.853.8/45.373.5/67.8

RPM outperforms PAEMA, DRE, and DKUA in both mAP and R@1 on every one of the seven targets under both training orders. Its Unseen-Avg exceeds PAEMA’s by 3.5%/4.4% in Order-1 and 3.0%/4.0% in Order-2. Thus, the improvements extend across the individual target datasets and are supported by their measured retrieval results.

(2) Regarding the choice of r=α=64r=\alpha=64. In Section E.4, we select the configuration by jointly considering Seen-Avg, Unseen-Avg, and memory cost. At r=α=64r=\alpha=64, RPM achieves Seen-Avg of 65.2/76.7 and Unseen-Avg of 73.3/67.6. Increasing the rank to 128 improves Seen-Avg by only 0.9%/0.9%, while reducing Unseen-Avg by 0.7%/1.5% and doubling the rank-dependent LoRA parameter storage. We therefore adopt 64 as a favorable performance–storage trade-off. The unseen-domain evaluation scores do inform this hyperparameter choice, although these domains are excluded from network training. The reported generalization results concern transfer to training-unseen domains under this configuration-selection procedure.

AuthorsOpenReview ↗

Response to Reviewer aA2h (2/3): Differences from InfLoRA/CoLoR and matched ReID comparisons

Thank you for highlighting InfLoRA and CoLoR and requesting clearer positioning and matched comparisons. Our response is provided below:

W2. Differences from InfLoRA and CoLoR.

Differences from InfLoRA and CoLoR. We conducted matched ReID experiments comparing our method with InfLoRA and CoLoR. RPM differs from both methods in how learned updates are written, retained, and used for each input. InfLoRA projects current-task input activations against historical subspaces, constructs a frozen input factor by SVD, and merges each learned update into the accumulated model. RPM instead applies SVD to projected current-domain second moments collected with all LoRA updates disabled, retains each learned update as a separate frozen domain memory, and reads these memories with input-dependent weights. Thus, the distinction concerns the construction and subsequent reuse of parameter memories; historical projection and a frozen SVD-designed input factor are shared components.

CoLoR independently trains dataset-specific LoRA experts and uses clustered features from the LoRA-disabled backbone to select the single expert associated with the nearest prototype. RPM combines history-constrained writing with soft multi-memory composition. Its maximum cosine affinity to each memory’s anchors is normalized across memories, and the resulting weights are combined with learned layer-wise amplitudes:

hl′=hl+∑i=1tgi(x)exp⁡(γil)Δil(hl)h_l'=h_l+\sum_{i=1}^{t}g_i(x)\exp(\gamma_i^l)\Delta_i^l(h_l).

This allows an image to draw on multiple domain memories without assigning it to a single training domain. The shared use of LoRA-disabled prototype features therefore does not make the two readout mechanisms equivalent.

Table 5 provides direct evidence for this readout choice within RPM. Soft reading achieves Seen 65.2/76.7 and Unseen 73.3/67.6, compared with hard prototype reading’s 61.0/72.3 and 52.8/45.2, respectively, in mAP/R@1. This is an internal ablation, not a matched evaluation of CoLoR.

For the matched comparison, we used the same ImageNet-1K-pretrained TransReID ViT-B/16 initialization, training-domain sequences, training budget, and evaluation splits. InfLoRA retained its historical-subspace construction and update merging; CoLoR retained independently trained experts and nearest-prototype expert selection. Both were adapted to ReID feature retrieval using identity classification and triplet objectives. Hyperparameters were selected using source-domain validation, and actual parameter/state budgets were recorded rather than assumed equal from rank alone. Both tables below use the same matched Ours reference runs. Results are reported as mAP/R@1.

MethodOrderSeen-AvgUnseen-Avg
InfLoRA157.6/70.269.1/62.1
CoLoR154.1/67.444.7/40.2
Ours165.2/76.773.3/67.6
InfLoRA257.0/69.267.5/60.9
CoLoR253.3/66.943.1/38.6
Ours264.5/75.773.5/67.8
AuthorsOpenReview ↗

Response to Reviewer aA2h (3/3): Unprojected-SVD ablation isolates historical projection

Thank you for suggesting the unprojected-SVD control to isolate the effect of historical projection. Our response is provided below:

W2. Isolating historical projection.

Isolating historical projection with unprojected SVD. We conducted a controlled ablation that removes only historical projection while retaining data-dependent SVD initialization and the frozen input factor. For the unprojected variant, we decomposed Ctl=UtlΣtl(Vtl)⊤C_t^l=U_t^l\Sigma_t^l(V_t^l)^\top and initialized Atl=(Ut,:,1:rl)⊤/3A_t^l=(U_{t,:,1:r}^l)^\top/\sqrt{3}. RPM applied the same operation to Pt−1lCtlP_{t-1}^lC_t^l. Both variants froze A, learned B, and retained the same rank, scale, statistics samples, historical-state construction, training budget, and reading protocol.

For the matched comparison, we used the same ImageNet-1K-pretrained TransReID ViT-B/16 initialization, training-domain sequences, training budget, and evaluation splits. Hyperparameters were selected using source-domain validation, and actual parameter/state budgets were recorded rather than assumed equal from rank alone.

Seen/Unseen scores are mAP/R@1; peak AF and BWT were computed separately for both metrics from actual stage-by-domain evaluations and are reported in percentage points. The projected-SVD rows use the same Ours runs as the method comparison above.

OrderInitializationSeen-AvgUnseen-AvgPeak AFBWT
1Unprojected SVD65.0/76.773.3/66.75.3/3.5−5.3/−3.4
1Projected SVD (Ours)65.2/76.773.3/67.64.5/2.9−4.4/−2.8
2Unprojected SVD64.1/75.373.3/67.06.8/5.4−6.8/−5.4
2Projected SVD (Ours)64.5/75.773.5/67.85.8/4.5−5.8/−4.5

This projected-versus-unprojected comparison isolates the incremental effect of historical projection within RPM without changing data dependence or input-factor trainability. Table 4’s existing random-versus-projected comparison does not separate these effects; its random fixed-A result remains an auxiliary reference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.