acceptodds
Under review as a conference paper at ICLR 2027

Not All Progress Transfers: Adaptation-State Reuse for Generalizable Object ReID

Abstract

Generalizable object re-identification aims to learn identity-discriminative representations that remain effective across heterogeneous object categories and visual distributions. Learning must reconcile fine-grained specialization with unknown matching requirements, although seen-category supervision gives no direct guidance on which adaptation changes to retain for transfer. We identify Progressive Feature Repurposing (PFR): structure beyond an early discriminative span gains seen-category retrieval utility while its category-readout accessibility declines modestly even after refitting. Restoring earlier structure benefits some transfer settings but harms others, with heterogeneous query-level responses: later progress does not uniformly supersede historical utility. Motivated by PFR, we propose ReState (Reusing Adaptation States), treating adaptation history as a reusable representational resource rather than committing to one selected state. During adaptation, ReState retains historical states and composes their feature residuals relative to a pretrained anchor through input- and dimension-conditioned routing. A zero-adaptation option retains access to anchor content, while optimization- and supervision-level regulation shapes alternatives for reuse. Across ten heterogeneous retrieval tasks, ReState demonstrates strong performance against existing ReID methods, with particularly large gains on multi-category benchmarks. It also improves mAP over a strong adaptation baseline on every evaluated category-shifted task.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generalizable Object Re-Identification, Cross-Category Generalization, Representation Adaptation

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_6mVYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates generalizable object re-identification (ReID), focusing on how representations learned during adaptation transfer to unseen object categories. The authors identify a phenomenon called Progressive Feature Repurposing (PFR), where continued adaptation improves discrimination on seen categories but does not necessarily improve transfer performance, while earlier representations may retain useful information. Based on this observation, they propose ReState, which consists of three components: (1) Trajectory Residual Routing (TRR), which combines historical adaptation states through input- and dimension-dependent routing; (2) Dual-Scale Trajectory Regularization (DTR), which regulates gradient updates and accumulated adaptation; and (3) Dual-Space Relation Perturbation (DSRP), which modifies classification and metric-learning supervision to reduce over-specialization to seen identities.

The method is evaluated on ten retrieval tasks covering different object categories and distribution shifts. The reported results show improvements over a strong adaptation baseline across all seven category-shifted tasks and competitive performance against existing ReID methods. Additional experiments analyze adaptation trajectories, historical feature restoration, component contributions, and routing strategies to support the proposed approach.

Critical strengths
  1. The paper provides an interesting analysis of how representations change during ReID adaptation. The results show that improvements on seen categories do not always translate into better performance on unseen categories, and that earlier representations can still be useful. The historical restoration experiments and query-level analysis provide evidence for this observation, giving a reasonable motivation for retaining intermediate adaptation states.
  2. The experimental evaluation covers a wide range of object categories and retrieval settings. ReState improves over the GVA baseline on all seven category-shifted tasks and achieves competitive results against existing ReID methods. The component ablations and comparisons with simpler routing strategies also help identify where the improvements come from. These results suggest that reusing adaptation history can be useful for cross-category generalization, even though the necessity of the full routing mechanism remains unclear.
Critical weaknesses
  1. The novelty of Progressive Feature Repurposing (PFR) is not sufficiently established. The observation that continued adaptation improves source-task performance while potentially reducing transferability is closely related to existing findings on feature distortion and over-specialization during fine-tuning. Although the authors provide additional analyses of feature subspaces and historical restoration, it remains unclear what fundamentally distinguishes PFR from these known phenomena. This weakens the significance of the paper's central motivation.
  2. The necessity of the proposed Trajectory Residual Routing (TRR) mechanism is questionable. Appendix Table 9 shows that TRR does not consistently outperform simpler alternatives, including static coordinate-wise routing and input-conditioned state-wise routing. For example, on University D→S, static coordinate-wise routing achieves 28.4 mAP, compared with 27.8 for TRR. Similar results appear on CUTE and ILIAS. These findings suggest that the additional complexity of combining input-dependent and dimension-wise routing may not be justified by its performance.
  3. The contribution of historical state reuse is not clearly separated from trajectory regularization. In Table 3, DTR alone provides larger improvements than TRR alone on several category-shifted tasks. For instance, on University D→S, DTR achieves 28.9 mAP, while TRR reaches 27.8. This raises the question of whether the main improvements come primarily from controlling the adaptation process rather than reusing historical representations. Although the full method performs well, the experiments do not fully establish the independent necessity of its components.
  4. The computational cost of ReState is not adequately evaluated. At inference, TRR requires extracting representations from multiple historical adaptation states and the pretrained anchor. Sharing the frozen backbone parameters reduces storage requirements but does not eliminate the cost of multiple forward passes. The paper does not provide measured inference latency, FLOPs, or a performance-cost comparison with single-state alternatives. Given that the routing improvements over simpler variants are relatively small, this additional inference overhead is important when assessing the practical value of the method.
Questions for authors
  1. How does Progressive Feature Repurposing (PFR) differ from previously studied feature distortion or loss of transferability during fine-tuning? What evidence supports treating PFR as a distinct phenomenon rather than another manifestation of over-specialization?
  2. In Appendix Table 9, TRR performs similarly to or worse than simpler routing strategies on several datasets. Could the authors explain why both input-dependent and dimension-wise routing are necessary? Under what conditions does their combination provide a clear advantage?
  3. Table 3 shows that DTR alone contributes more than TRR alone on several category-shifted tasks. Could the authors provide a clearer analysis of how much the final improvement comes from historical state reuse versus trajectory regularization? In particular, how does TRR perform when applied to states trained with DTR but without DSRP?
  4. Could the authors report the actual inference latency, FLOPs, and memory overhead of ReState compared with a single adapted encoder? How does retrieval performance change when fewer historical states are retained? This would help determine whether the additional computation is justified by the reported gains.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 6mVY (Part 1)

We thank the reviewer for their comments and suggestions.

1. How does PFR substantively differ from existing studies of fine-tuning distortion? (W1/Q1)

Response: PFR reveals a mismatch across training stages: while source identity discrimination continues to improve, earlier features can still improve the final model's instance retrieval on unseen categories. Existing studies of fine-tuning distortion [1,2] and OOD trajectories [3] do not directly answer whether these historical features can improve identity matching after training is complete. We examine this through a fixed early discriminative subspace, refitted category readouts, and restoration of early features.

Diagnostic taskFinal-model mAPmAP after restoring early featuresGain
PetFace41.6144.67+3.06
University D→S24.4929.28+4.79
CUTE50.0451.88+1.84
ILIAS36.0338.44+2.41

We consistently report four representative tasks with established restoration results, covering animal, cross-view, and multi-category retrieval. The complete positive and negative restoration effects and random-restoration controls are in Appendix A.6, Table 6; category readouts are in Appendix A.5, Table 5. ReState's design advantage is to retain and reuse these historical states without having to choose a stopping point in advance for unseen targets.

2. Does input-dependent, dimension-wise TRR outperform simple combinations? (W2/Q2)

Response: TRR computes historical-state combination weights for each feature dimension based on changes in the input image's features; with the same state pool, it outperforms fixed feature averaging using the same number of encoder passes. The added same-trajectory experiment shows all three category-aligned tasks, University D→S, and all three multi-category tasks:

Task (mAP)Equal-weight feature averageTRR
MSMT1725.8425.45
VehicleID55.9155.87
ShipReID-24004.484.45
University D→S27.3627.79
PetFace42.6743.26
CUTE50.8751.21
ILIAS37.6038.15

Equal-weight averaging is slightly better on the three category-aligned tasks; TRR leads on the four displayed unseen-category tasks and on all seven unseen-category tasks in the complete evaluation. This demonstrates the value of learned combination weights for category transfer; the advantage cannot be explained solely by additional encoder passes. Other simplified routing forms are reported in Appendix B.7, Table 9.

3. Given DTR's gains, how much additional contribution does TRR provide? (W3/Q3)

Response: TRR combined with DTR outperforms DTR alone on all five tasks in the component experiment. DTR changes state formation through gradient reversal and LoRA shrinkage, while TRR reuses features from these stages, making their roles complementary.

Task (mAP)TRRDTRTRR+DTR
ShipReID-24004.54.65.2
University D→S27.828.931.5
PetFace43.344.244.7
CUTE51.253.153.3
ILIAS38.239.942.7

We report all five component-experiment tasks from Table 3 and Appendix B.7, Table 10; this is a different experiment from the repeated training across ten tasks. State formation and reuse can jointly improve unseen-category retrieval.

4. Does multi-state inference incur excessive cost for limited gains? (W4/Q4)

Response: The actual overhead can be substantially reduced through fewer states and multi-GPU parallelism; full ReState raises the ten-task mean mAP from GVA's 26.65 to 29.08, with gains on all seven unseen-category tasks. State features can be encoded independently and then combined. Measurements on A800 GPUs with batch=512 are:

ConfigurationGPUGFLOPs/imageFPSBatch latency msMaximum per-GPU peak GiB
OG-ReID (CVPR 2026)146.583338.081514.4224.310
ReState structure, 10 states1696.437161.663167.068.547
ReState structure, 5 states1377.915300.051706.366.019
ReState structure, 5 states2377.915581.54880.424.527

On one GPU, five-state throughput is only 11.2% lower than OG-ReID (CVPR 2026), with 12.7% higher batch latency and 75.2% lower peak memory; two GPUs further increase throughput. The default ten-state configuration is not the lowest-cost option. TRR also outperforms equal-weight averaging with the same state pool (Question 2), showing that effective feature combination contributes to the accuracy gains. Future work can further investigate which historical states are worth retaining.

AuthorsOpenReview ↗

Response to Reviewer 6mVY (Part 2)

5. Does retaining fewer historical states reduce retrieval accuracy? (Q4)

Response: In the existing cache-interval experiment, fewer states instead improve mean mAP while reducing measured latency.

Cache intervalState countFour-task mean mAPSingle-image latency ms
500 (default)1040.1092.61
600840.4876.94
800640.8162.20
900541.1254.96

This experiment consistently evaluates PetFace, University D→S, CUTE, and ILIAS. The mAP values come from independent GVA+TRR training runs with different cache intervals; latency comes from structure-level measurements on one A800 with batch=1 (see Appendix B.6, Fig. 6 for the parameter sweep). Fixed intervals do not guarantee selection of the most suitable historical states, and improved state selection may further increase efficiency.

References

[1] Kumar, A., Raghunathan, A., Jones, R., Ma, T. and Liang, P. "Fine-Tuning Can Distort Pretrained Features and Underperform Out-of-Distribution." ICLR, 2022.

[2] Mukhoti, J., Gal, Y., Torr, P.H. and Dokania, P.K. "Fine-Tuning Can Cripple Your Foundation Model; Preserving Features May Be the Solution." Transactions on Machine Learning Research (TMLR), 2024.

[3] Andreassen, A., Bahri, Y., Neyshabur, B. and Roelofs, R. "The Evolution of Out-of-Distribution Robustness Throughout Fine-Tuning." Transactions on Machine Learning Research (TMLR), 2022.

Reviewer_S2AoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies source-only object ReID using DINOv2 ViT-B/14 with LoRA, trained on five categories and evaluated on ten tasks. Its PFR analysis shows that early feature structure retains transfer utility even as adaptation improves source retrieval. ReState combines snapshot residuals through input and dimension-conditioned routing (TRR), with occasional gradient reversal/adapter shrinkage (DTR) and label perturbation (DSRP). In the experiments, ReState outperforms the compared ReID methods on eight of the ten tasks and its own baseline, GVA, on all seven tasks with a category shift.

Critical strengths

S1) The evaluation is extensive. Ten heterogeneous target tasks in three transfer regimes, three metrics and detailed ablations in the appendix.

S2) The reported gains over the matched baseline GVA are consistent on all seven category-shifted tasks, with defaults stated to be selected on source data only.

Critical weaknesses

W1) The ablations do not show that the full design is needed. LUR alone, via a single encoder pass, is within two mAP points of full ReState on six tasks and better on three, while ReState evaluates eleven encoder states per query. FLOPs and latency are not reported. The simpler routers in Table 9 perform on par with the proposed one. DSRP gains at most one point over random relabeling with the same quotas (Table 11), and adding it to TRR+DTR lowers ILIAS and ShipReID. This configuration is left out of Table 3, as are the two aligned tasks where ReState is below GVA. Most of these differences are below one mAP point and come from a single seed, with bootstrap intervals covering only evaluation noise.

W2) Simple baselines are missing. Snapshot Ensembles, Model Soups and WiSE-FT are discussed in Section 2 but never run. The potential options are a weight average of the LoRA updates, and scaling the final update toward the pretrained anchor, at the cost of one encoder pass. Single-pass regularizers toward the pretrained model, L2-SP (cited but not run) and LDIFS [4], address the same problem. [3] studies the same question on fine-grained domains and is not cited, shows that they recover most of the lost out-of-domain retrieval, although with external validation data. Finally, as the task is cross-category instance-level retrieval, the frozen DINOv2 and universal embeddings [1, 2] should be evaluated on the ten targets.

[1] Towards Universal Image Embeddings: A Large-Scale Dataset and Challenge for Generic Image Representations, Ypsilantis et al., ICCV 2023

[2] Unicom: Universal and Compact Representation Learning for Image Retrieval, An et al., ICLR 2023

[3] Infusing Fine-Grained Visual Knowledge to Vision-Language Models, Ypsilantis et al., ICCVW 2025

[4] Fine-Tuning Can Cripple Your Foundation Model; Preserving Features May Be the Solution, Mukhoti et al., TMLR 2024

W3) The PFR analysis adds little beyond the cited fine-tuning literature and does not motivate the design. Matched-random restoration is as good as or better than restoring the fitted complement on several targets, as acknowledged, so the complement plays no specific role. The category-readout drop is below one percentage point on thirty held-out identities, and the early gains are not separated from the decaying learning rate. What remains, from one trajectory and backbone, is that earlier checkpoints transfer better to some targets, which is known, and nothing favors the learned router over the averaging controls of W2.

W4) The writing obscures simple mechanisms. Coined terms (PFR, TRR, DTR, DSRP, LUR, IR, GVA, Shared Anchor, Moving State) replace plain statements, and the configuration values appear only in the appendix. Few examples of the cryptic text:

  • "Local Update Reversal occasionally opposes the current loss-gradient contribution" means with probability 0.05 the gradient is multiplied by -4.
  • "Initialization Retraction contracts accumulated adaptation toward the pretrained reference" means shrink the LoRA weights toward zero.
  • "Dual-Space Relation Perturbation selectively varies supervision" means deliberately change a few training labels.
  • "Progressive Feature Repurposing: later progress does not uniformly supersede historical utility" means earlier checkpoints transfer better to some targets.

W5) The main comparison does not separate ReState from its own baseline. GVA appears in the ablations but not in Table 2, although it alone exceeds every external method on five of the ten tasks, most of the margin over prior work is already in the baseline. The backbones and training setups of the compared methods are not given. BAU and ADSR obtain only 1.8 and 0.7 mAP on MSMT17, although Market-1501 is among the training sources. For context, ResNet-50-based QAConv-GS and TransMatcher report 17.2 and 18.4 mAP after training on Market-1501 alone. The paper should state whether these are copied or re-run, and with which configurations.

Questions for authors

Q1) Can you add GVA, the averaging controls of W3 and L2-SP/LDIFS to Table 2, with source-selected coefficients and query-encoding cost, together with frozen DINOv2, and a universal embedding representative? Potentially also add the simple single-pass baseline of regularization only as in [3].

Q2) Do the gains hold over at least three seeds for GVA, LUR, TRR+DTR and full ReState?

Q3) Which source-data split, metric and aggregation rule selected the cache interval, amplification and quota in B.6. Was a leave-one-source-out strategy used?

5 comments
AuthorsOpenReview ↗

Response to Reviewer S2Ao (Part 1)

We thank the reviewer for their comments and suggestions.

1. LUR is already strong with a single encoder pass; is multi-state inference worthwhile? (W1/Q1)

Response: LUR demonstrates the effectiveness of our training strategy with a single encoder; ReState additionally retains historical states, outperforms GVA in unseen-category retrieval, and can reduce inference overhead through parallel execution. The three-seed mean mAP results below cover all three category-aligned tasks, University D→S, and all three multi-category tasks:

Task (mAP)GVALURReState
MSMT1725.2625.9023.46
VehicleID55.6456.3355.69
ShipReID-24004.275.074.96
University D→S24.7132.1832.51
PetFace41.6944.8045.74
CUTE48.7351.1552.68
ILIAS35.6939.7440.46

Compared with GVA, full ReState improves mAP on all seven unseen-category tasks; compared with LUR, it still provides gains on tasks such as PetFace and CUTE. On A800 GPUs with batch=512, the five-state structure reaches 300.05 FPS on one GPU, close to OG-ReID's (CVPR 2026) 338.08 FPS, and 581.54 FPS on two GPUs. TRR outperforms equal-weight averaging with the same number of states (Questions 2 and 6), showing that the feature-combination mechanism also contributes to its effectiveness.

2. Given the similar performance of simpler routers, what is TRR's role? (W1)

Response: With the same historical states and forward computation, TRR improves unseen-category retrieval through combination weights that depend on the input and feature dimension. The added controls from the same training trajectory consistently report seven tasks:

Task (mAP)Equal-weight feature averageTRR
MSMT1725.8425.45
VehicleID55.9155.87
ShipReID-24004.484.45
University D→S27.3627.79
PetFace42.6743.26
CUTE50.8751.21
ILIAS37.6038.15

TRR outperforms equal-weight averaging on all seven unseen-category tasks, while equal-weight averaging is slightly better on the three category-aligned tasks. Other routing forms and more detailed controls are reported in Appendix B.7, Table 9. This shows that TRR's main value lies in category transfer, without requiring more states.

3. Does DSRP outperform random relabeling, and is it effective when added to TRR+DTR? (W1)

Response: DSRP uses categories, prediction scores, and feature distances to select which supervision relationships to modify; it outperforms random changes with the same quota on the four cross-view and multi-category tasks below.

Task (mAP)Random relabelingDSRP
University D→S25.926.9
University S→D16.517.4
CUTE51.251.3
ILIAS36.837.6

This experiment specifically examines label-modification rules, with a task scope different from the standard ReState comparisons; the complete comparison is in Appendix B.7, Table 11. After adding DSRP, ReState's ten-task mean mAP improves over TRR+DTR in all three independent seeds. Thus, selectively adjusting source supervision can further improve overall results on top of historical reuse and DTR.

4. Why were TRR+DTR and the category-aligned tasks not shown together in the main text? (W1)

Response: Given the main-text space limit, Table 3 selected representative tasks from different categories; the original TRR+DTR component results were already provided in Appendix B.7, Table 10. To fully address the concern about category-aligned tasks, we use three-seed means here and consistently show all three category-aligned tasks and four representative unseen-category tasks:

Task (mAP)GVATRR+DTRReState
MSMT1725.2622.3223.46
VehicleID55.6454.9055.69
ShipReID-24004.275.144.96
University D→S24.7131.8932.51
PetFace41.6944.7645.74
CUTE48.7352.2852.68
ILIAS35.6941.6240.46

ReState outperforms GVA on all seven unseen-category tasks; TRR+DTR's performance on tasks such as ShipReID-2400 and ILIAS also demonstrates the complementarity of historical reuse and DTR. The category-aligned results are presented alongside them to make the scope of the gains clear.

AuthorsOpenReview ↗

Response to Reviewer S2Ao (Part 2)

5. Do three-seed training results demonstrate reproducibility? (W1/Q2)

Response: After three independent training runs, ReState retains the highest mean mAP on University D→S, PetFace, and CUTE, while TRR+DTR is higher on ILIAS; full ReState still has the highest ten-task mean mAP among the four configurations. We show one cross-view task and all three unseen multi-category tasks, reporting mean ± standard deviation across training runs:

Task (mAP)GVALURTRR+DTRReState
University D→S24.71 ± 0.8632.18 ± 1.1231.89 ± 0.3932.51 ± 0.27
PetFace41.69 ± 0.3344.80 ± 0.3444.76 ± 0.4245.74 ± 0.39
CUTE48.73 ± 1.1251.15 ± 1.0752.28 ± 1.3652.68 ± 0.67
ILIAS35.69 ± 0.4039.74 ± 0.4641.62 ± 0.9640.46 ± 0.56

These independently trained results support ReState's stable gains in cross-category retrieval; unlike the original evaluation resampling for a fixed model, the three-seed experiment directly examines training randomness.

6. Can simple parameter or feature averaging replace TRR? (W2/W3/Q1)

Response: With the same training trajectory and the same historical states, TRR outperforms equal-weight feature averaging on all seven unseen-category tasks without requiring additional state encoding. The added matched controls are:

Task (mAP)Final stateLoRA-parameter averageEqual-weight feature averageTRR
MSMT1724.8825.6225.8425.45
VehicleID55.6455.9055.9155.87
ShipReID-24004.124.344.484.45
University D→S22.9826.9327.3627.79
PetFace40.5942.5242.6743.26
CUTE48.9050.7050.8751.21
ILIAS34.8137.4637.6038.15

Historical states are useful in themselves, and TRR further learns which stages' features suit the current image. The table also retains the three category-aligned tasks. Existing routing-form ablations are in Appendix B.7, Table 9.

7. Can single-encoder regularization methods L2-SP and LDIFS replace historical reuse? (W2/Q1)

Response: ReState has higher seven-task mean mAP than L2-SP [1] and LDIFS [2], indicating that preserving only the parameters or features of a single final model is insufficient to fully use the discriminative capabilities from different adaptation stages. We evaluated L2-SP [1] and LDIFS [2] in a source-only setting; the table below follows the fixed seven-task presentation.

Task (mAP)L2-SP [1]LDIFS [2]ReState
MSMT175.422.5623.46
VehicleID54.6741.6755.69
ShipReID-24002.881.334.96
University D→S35.1725.6832.51
PetFace27.2122.1445.74
CUTE34.7925.7352.68
ILIAS44.7225.9540.46
Seven-task mean29.2720.7236.50

ReState outperforms LDIFS on all seven tasks and L2-SP on five. Its advantage is consistent with TRR's use of multiple historical states and DTR and DSRP's regulation of state formation.

8. What are the novelty and gains relative to the fine-grained feature-preservation results in [1]? (W2)

Response: The single-model preservation setting in [1] uses only the final representation; ReState retains historical features and combines them according to the input, achieving higher seven-task mean mAP under the same source-only setting. The L2-SP [1] comparison uses the same seven tasks:

Task (mAP)L2-SP [1]ReState
MSMT175.4223.46
VehicleID54.6755.69
ShipReID-24002.884.96
University D→S35.1732.51
PetFace27.2145.74
CUTE34.7952.68
ILIAS44.7240.46
Seven-task mean29.2736.50

ReState leads L2-SP on five of the seven tasks. For ReID with unknown target categories, retaining states trained to different degrees on source identities enables retrieval to use features that a single final model does not fully exploit.

AuthorsOpenReview ↗

Response to Reviewer S2Ao (Part 3)

9. Are frozen DINOv2 and existing general-purpose representations sufficient for cross-category ReID? (W2/Q1)

Response: ReState outperforms frozen DINOv2 on all seven tasks and has higher seven-task mean mAP than L2-SP and LDIFS. We report results on exactly the same seven tasks as in Questions 7 and 8:

Task (mAP)Frozen DINOv2L2-SP [1]LDIFS [2]ReState
MSMT170.335.422.5623.46
VehicleID32.8454.6741.6755.69
ShipReID-24001.242.881.334.96
University D→S23.2135.1725.6832.51
PetFace44.4027.2122.1445.74
CUTE51.6534.7925.7352.68
ILIAS33.3444.7225.9540.46
Seven-task mean26.7229.2720.7236.50

These results show that general pretrained features alone do not fully meet the requirements of fine-grained identity matching; ReState obtains stronger cross-category retrieval representations through source identity adaptation and historical-state reuse.

10. Does PFR merely observe that some early checkpoints perform better? (W3)

Response: PFR's key experiment restores earlier features into the final model to directly test whether they can still improve final retrieval, rather than merely comparing scores from early and late checkpoints. We use the same four restoration-diagnostic tasks as in our response to 6mVY:

Task (mAP)Final modelAfter restoring early features
PetFace41.6144.67
University D→S24.4929.28
CUTE50.0451.88
ILIAS36.0338.44

The complete positive and negative effects and random-restoration results are in Appendix A.6, Table 6; the fixed subspace and category readouts are in Fig. 2 and Appendix A.4–A.5. Historical features can continue to provide practical retrieval value after training is complete, directly supporting ReState's state-reuse design.

11. If random-subspace restoration is also effective, why analyze a fitted complement? (W3)

Response: The fitted complement provides a fixed reference for observing how feature uses change during training; useful information in historical states is not limited to this subspace. Matched-random-subspace restoration is similarly effective on some targets (Appendix A.6, Table 6), indicating that useful features are distributed across multiple directions. ReState uses complete historical representations and requires no fitting or selection of a particular subspace when deployed to unseen categories.

12. Given the small category-readout decline, is PFR sufficiently meaningful? (W3)

Response: Category classification differs from ranking identities within the same category; PFR's most direct evidence is that restoring early features can improve final ReID. Category readouts provide supplementary evidence of functional changes:

DiagnosticAccuracy change / percentage points95% evaluation interval
Fixed early classifier−3.27[−4.16, −2.39]
Refitted complement classifier−0.78[−1.16, −0.44]

The readout experiment is in Appendix A.5, Table 5; actual retrieval restoration is in Appendix A.6, Table 6. The method is motivated by real retrieval gains from earlier features and does not require a large decline in category accuracy.

13. Are the early gains caused by the decaying learning rate? (W3)

Response: The benefit of historical reuse can be observed directly under exactly the same learning-rate trajectory. The same-trajectory controls on the fixed seven tasks in Question 6 show that changing only how stored features are used can outperform the final state on unseen-category tasks. A decaying learning rate may affect how quickly different training stages change, but it does not eliminate the practical value of historical states already formed for final retrieval. The diagnostic trajectory settings are in Appendix A.1.

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

The manuscript studies generalization in object re-identification (ReID) under source-only adaptation. It introduces the notion of Progressive Feature Repurposing (PFR): continued identity-discriminative adaptation improves source retrieval, but the measured functions of the representation change, and earlier feature structure is not uniformly superseded by later checkpoints. The authors support PFR with diagnostics on fixed subspaces, category readout probes, restoration interventions, and query-level response distributions. Motivated by these observations, they propose ReState, which couples adaptation-state formation with state reuse. ReState contains three components: Trajectory Residual Routing (TRR), which retains periodic LoRA snapshots and composes their residuals relative to a pretrained anchor with input- and coordinate-wise routing; Dual-Scale Trajectory Regularization (DTR), which regulates update accumulation through local update reversal and initialization retraction; and Dual-Space Relation Perturbation (DSRP), which selectively perturbs classification and metric supervision. Experiments use DINOv2 ViT-B/14 with LoRA on five source categories and evaluate on ten tasks covering category-aligned and category-shifted retrieval. The authors report improvements over a strong GVA baseline on all seven category-shifted tasks and competitive complete-system results across 111 dataset-local categories.

Critical strengths
  1. The diagnostic analysis is more nuanced than a simple early-stopping argument. PFR is supported by three complementary lines of evidence: continued development of complementary discriminative structure, reduced category-readout accessibility, and heterogeneous restoration responses across settings and queries. The finding that dataset averages conceal opposing query-level responses is particularly useful.
  2. The problem formulation couples state formation and state reuse. Rather than treating checkpoints as an ensemble or selecting a single adaptation depth, ReState learns how to compose adaptation residuals while also regulating the optimization and supervision that produce reusable states. This is a coherent framing.
  3. TRR is a reasonable and flexible design. Residualizing states around a pretrained anchor, then routing per input and per coordinate, allows different feature dimensions to draw on different adaptation stages. The chronology-aware regularizer is a sensible safeguard against premature collapse toward late states.
Critical weaknesses
  1. Limited backbone and source-category diversity. The main experiments use a single DINOv2 ViT-B/14 backbone with LoRA and five source categories. Although evaluation covers 111 dataset-local categories, it is unclear whether the conclusions transfer to other pretrained encoders, larger backbones, or broader source category sets.
  2. Many components and hyperparameters increase tuning cost. ReState involves TRR, DTR (LUR and IR), and DSRP, with cache interval, amplification factor, perturbation quota, routing temperature, and chronology weight. The sensitivity analyses are useful, but the combined tuning space is large, and the paper does not provide a principled procedure for selecting these hyperparameters in new deployments.
  3. Inference cost of multi-state encoding is not fully characterized. Shared backbone weights reduce storage, but each retained state still requires a forward pass, and query-side computation grows with the number of states. The appendix acknowledges this, but the paper does not report end-to-end latency, throughput, or memory as a function of pool size, nor does it compare against single-model combination methods under matched inference budgets.
Questions for authors
  1. Add at least one additional pretrained encoder or backbone scale, and if feasible one additional source-category grouping. If this is infeasible, explicitly bound the generality claim to the current DINOv2 + LoRA + five-source setting.
  2. Provide latency, throughput, and memory as a function of the number of retained states, and compare against Model Soups, WiSE-FT, VRF, DaWin, or Twin-Merging under matched inference budgets where possible.
2 comments
AuthorsOpenReview ↗

Response to Reviewer RjJY (Part 1)

We thank the reviewer for their comments and suggestions.

1. Can ReState be applied to visual backbones of different sizes? (W1/Q1)

Response: ReState has been evaluated with all four DINOv2 sizes—S, B, L, and g—and benefits from stronger backbones in cross-view and unseen multi-category retrieval. We show all three category-aligned tasks, University D→S, and all three multi-category tasks below, from the separate backbone-size experiment:

Task (mAP)DINOv2-SDINOv2-BDINOv2-LDINOv2-g
MSMT1718.4123.2426.1921.82
VehicleID53.9955.5656.7155.37
ShipReID-24003.644.814.815.74
University D→S24.8032.6743.7049.91
PetFace38.0045.6450.8950.98
CUTE41.1753.3257.7254.58
ILIAS28.9941.1049.6249.95

ReState can use historical states from adaptation across multiple backbone sizes. Training and data settings are given in Appendix B.4–B.5.

2. Does ReState retain its advantage when the source training categories change? (W1/Q1)

Response: After changing the composition of the five source training datasets, ReState still outperforms existing methods on all five target tasks, showing that its cross-category retrieval advantage does not depend on a particular combination of source categories. The new training sources are Market-1501, ATRW, ELPephants, ShipReID-2400, and DenseUAV.

Method (mAP)VesselReIDVehicleIDCUTEILIASPetFace
TransReID (ICCV 2021)12.7029.1020.007.2026.70
PHA (CVPR 2023)13.9028.2021.408.5029.10
CLIP-ReID (AAAI 2023)18.3042.9022.8016.0024.70
PAT (ICCV 2023)19.9036.8028.4011.3036.00
PromptSG (CVPR 2024)12.6027.9018.907.7021.20
CLIP-FGDI (TIFS 2025)16.8136.9225.7410.5119.95
VICP (ICCV 2025)3.8218.3015.695.7222.17
OG-ReID (CVPR 2026)23.0740.6030.3223.6239.63
ReState32.0146.5150.3933.9145.83

ReState achieves the highest mAP on all five tasks, with particularly clear advantages on multi-category retrieval tasks such as CUTE and ILIAS. This benefits from ReState's ability to retain identity-discriminative features formed at different training stages and combine them adaptively according to the input, accommodating the retrieval requirements of different object categories.

The table includes all five target tasks in this new-source experiment. Dataset protocols are given in Appendix B.5, Table 8.

3. How should the default parameters be selected for a new target? (W2)

Response: New targets directly use the parameters determined on the source side: a cache interval of 500 steps, DTR amplification of 4, and a DSRP modification quota of 4, without requiring target identity labels. Definitions are provided in Appendix B.1–B.4, and parameter sensitivity is examined in Appendix B.6, Fig. 6. When deployment budgets differ, the cache interval is the primary way to control the number of historical states; the other parameters can retain their default values. We will publicly release our complete code and experimental configurations to facilitate reproducibility.

4. Does increasing the number of states make inference cost unmanageable? (W3/Q2)

Response: Inference cost can be controlled through the number of retained states; the five-state configuration already has measured latency close to OG-ReID (CVPR 2026), with lower memory usage. Measurements on one A800 with batch=512 are:

ConfigurationGFLOPs/imageFPSBatch latency msPeak GPU memory GiB
OG-ReID (CVPR 2026)46.583338.081514.4224.310
ReState, 10 states696.437161.663167.068.547
ReState, 8 states569.028198.172583.587.531
ReState, 5 states377.915300.051706.366.019

Compared with the default ten-state configuration, five states reduce FLOPs by approximately 45.7% and increase throughput by about 1.86×; two A800 GPUs reach 581.54 FPS. In the cache experiment on the fixed set of PetFace, University D→S, CUTE, and ILIAS, five states yield higher mean mAP than ten states (Appendix B.6, Fig. 6), showing that accuracy and cost can be improved together.

AuthorsOpenReview ↗

Response to Reviewer RjJY (Part 2)

5. Is inference overhead reasonable compared with other model-combination methods? (W3/Q2)

Response: ReState's multi-state computation can be substantially accelerated by retaining fewer cached states and executing them in parallel; the five-state structure achieves practical speed close to OG-ReID (CVPR 2026), with lower peak memory. Single-GPU measurements at batch=512 are:

Method / configurationFPSBatch latency msPeak GPU memory GiB
Model Soups (ICML 2022)8330.5261.460.912
WiSE-FT (CVPR 2022)8427.3260.760.911
VRF (NIPS 2024) (N=100,000)500.031023.942.808
DaWin (ICLR 2025) (BMM-3)2249.15227.641.090
Twin-Merging (NIPS 2024)60.088522.522.248
OG-ReID (CVPR 2026)338.081514.4224.310
ReState, 5 states300.051706.366.019

The gains from multi-state inference do not come solely from additional encoder passes: with exactly the same state pool and number of encoder passes, TRR outperforms equal-weight feature averaging on all seven unseen-category tasks, and combining it with DTR brings further improvements (Table 3; Appendix B.7, Tables 9–10). Different states can be encoded on separate GPUs, and future work can further improve the selection of cached states.

Comments on the paper

1 comment
Area_Chair_PgCmOpenReview ↗

Urgent Reminder: Reviewer Discussion Deadline

Dear Reviewers,

This is a friendly but urgent reminder that the discussion period will end on November 18, 2026. Please carefully review the authors' rebuttals and actively participate in the ongoing discussions. If you have not yet responded to the authors or initiated a discussion, please do so as soon as possible. In particular, reviewers are expected to address the authors' clarifications and responses, provide constructive follow-up feedback, and update their assessments and scores when appropriate. Please ensure that your final evaluations reflect your independent judgment and take into account the information provided during the rebuttal and discussion process. Active and timely participation is essential to ensuring a fair, thorough, and constructive review process.

Thank you for your time, effort, and cooperation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.