acceptodds
Under review as a conference paper at ICLR 2027

RAPR: Reliability-Adaptive Prompt Routing for Aerial-Ground Person Re-Identification

Abstract

Aerial-ground person re-identification matches pedestrians across drone and ground cameras, whose radically different imaging geometries alter body-part visibility, scale, and context. Consequently, local cues can range from identity-bearing garment details to view-induced clutter and missing regions, making their reliability highly uneven. We present Reliability-Adaptive Prompt Routing (RAPR), a Transformer framework that uses reliability to guide prompt adaptation and local aggregation. RAPR derives a view-reduced identity anchor and dynamically initializes prompts from a shared prompt and an input-conditioned prompt bank. It then selects anchor-consistent local evidence and routes global, reliable-local, and residual view-rejection contexts to refine the prompts. A local refiner separately estimates token identity relevance and view bias to calibrate local memory and guide bidirectional prompt–local interaction. The final descriptor combines global and refined local features. RAPR outperforms multiple state-of-the-art methods across evaluation protocols on several AG-ReID benchmarks. Code and model weights will be released on GitHub upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aerial-ground vision, Person re-identification, Prompt learning, Vision transformer, Cross-view retrieval

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JMNSOpenReview ↗
Rating
3
Confidence
3
Informativeness
3(+1)
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper presents RAPR, a Transformer-based framework for aerial–ground person re-identification that uses identity- and view-conditioned local evidence to guide prompt adaptation and feature aggregation. The encoder constructs a view-reduced identity anchor from global and view tokens. RAPR dynamically initializes prompts, selects local tokens using relative identity–view similarity, and adaptively routes complementary contexts to refine the prompts. A subsequent local refinement module learns identity-relevance and view-bias gates to calibrate patch features and guide bidirectional prompt–local interaction. Global and refined local descriptors are concatenated for retrieval.

Critical strengths

1.The paper is clearly presented and well motivated by a meaningful challenge in aerial–ground person re-identification: local cues vary substantially in visibility and usefulness across viewpoints. It clearly explains how unreliable local evidence can affect prompt adaptation and feature aggregation, providing a coherent rationale for the proposed two-stage design.

2.The framework learns local selection and refinement using image-level identity labels and coarse view labels, without requiring patch-level reliability annotations or explicit cross-view correspondences. Each image is also processed independently at inference. These requirements make the approach applicable to existing aerial–ground ReID datasets and avoid additional annotation and input-pairing burdens, supporting its practical usefulness.

Critical weaknesses

1.Sections 4.1 and A.2 state that baseline results are taken from the cited papers unless explicitly marked as reproduced, and SeCap has no reproduction marker in Table 8. However, its reported LAGPeR results differ from the cited publication. For A→G, Table 8 lists SeCap as 40.51/29.59 in Rank-1/mAP, whereas SeCap’s original paper, Table 3, reports 41.79/30.37, exceeding RAPR’s 41.33/29.93. The original SeCap also reports 24.39/19.24 for G→(A+G), above RAPR’s 23.54/18.96. These discrepancies change the rankings and undermine the claim that RAPR leads every metric on LAGPeR. The authors should clarify the source and evaluation settings of these numbers and provide reconciled comparisons.

2.The paper states that it follows the official dataset splits, but the partitioning and validation procedure could be described more explicitly. In particular, Appendix A.3.3 notes that the benchmarks lack official validation splits without specifying whether an independent validation subset was constructed. Please clarify the training, validation, and test partitions, including the numbers of identities and images where applicable, their identity-level separation, and the data used for hyperparameter selection. These details would improve the clarity and reproducibility of the experimental protocol.

3.Table 9 attributes successive improvements to adding the reliable local context (c_2), adding (c_3), and learning the routing weights. However, its first row is explicitly the VDIA-only configuration, which excludes DPI, RAR, and VSLR in Table 4, rather than the complete architecture using only (c_1). The comparison therefore does not establish that the remaining prompt and local processing components are held fixed when (c_2) is introduced. Consequently, the increase in CARGO ALL mAP from 61.20 to 65.23 cannot be uniquely attributed to (c_2). A clearer comparison should retain the same architecture, losses, and descriptor construction while varying only the available contexts.

Questions for authors

1.Could you clarify the source and evaluation settings of the SeCap results in Table 8 and explain their differences from the original publication?

2.Could you specify the training, validation, and test splits, including identity/image counts, identity-level separation, and the data used for hyperparameter selection?

3.Could you add a (c_1)-only baseline within the complete architecture and vary only the contexts, keeping all other components, losses, and descriptor construction fixed?

Reviewer_y48oOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RAPR addresses aerial-ground person re-identification through reliability-guided prompt adaptation and local aggregation. It combines identity and view references, sparse token selection, and dense local refinement. Experiments cover CARGO, AG-ReID.v2, and LAGPeR.

Critical strengths

1.Selecting evidence before prompt updating and retaining broader coverage during local aggregation is a coherent, task-motivated design.

2.The overall experiments and component studies support the method’s usefulness, with clear improvements over the controlled ViT baseline on CARGO.

3.Neither paired cross-view inputs nor patch annotations are required, and the additional computational cost over SeCap is modest.

Critical weaknesses

1.The novelty needs a stronger defense. View decoupling, adaptive prompts and bidirectional local refinement are already central to SeCap, while ViSA also explores adaptive token generation. RAPR adds a specific reliability criterion within this established pipeline. The rebuttal needs to explain why this amounts to a substantial methodological contribution beyond another refinement of prompt-based aggregation.

2.Why should view affinity count against identity reliability? View-dependent clothing and body cues can still be discriminative. The authors appropriately acknowledge that the anchor and gates have no guaranteed disentanglement or calibrated probability interpretation. However, the central reliability argument still needs justification. Please explain the conditions under which this criterion is reasonable and how useful view-correlated evidence is protected.

3.The recovery argument is incomplete. RAR selection, VSLR gating and local-to-anchor alignment all rely on the same global reference (Eqs. 4, 11 and 16). Dense refinement restores token coverage, but this alone does not explain recovery from a poor anchor. The rebuttal should clarify how the local branch preserves complementary identity evidence when the anchor is misleading, instead of reinforcing its errors.

Questions for authors

I currently lean toward weak accept because the design is coherent and the overall experiments support its usefulness. If the rebuttal does not adequately address the novelty and central mechanistic concerns in weaknesss, I will lower my score.

4 comments
AuthorsOpenReview ↗

Response to Reviewer y48o

Thank you for your careful review and Weak Accept assessment. We appreciate your recognition of our design and overall experimental results. Below, we address your concerns about novelty, view-correlated cues, and anchor dependence through revised explanations and additional analyses.

Response to Weakness 1: Novelty relative to SeCap and ViSA

We agree that view decoupling, adaptive prompts, and bidirectional local interaction have precedents. RAPR builds on these ingredients and introduces explicit decisions about which local evidence updates the prompt and how local memory is calibrated before descriptor aggregation.

MethodPrompt/token adaptation and local aggregation
SeCapRecalibrates learnable prompts using the view-decoupled global representation, followed by bidirectional prompt–local refinement.
ViSAGenerates semantic tokens through shared and view-specific experts, then uses Top-k local selection and dual-branch graph reasoning to fuse local features.
RAPRSelects local evidence using its relative affinity to identity and view references before it updates the prompt, adaptively routes three candidate contexts, and applies learned dense token calibration before local aggregation.

In particular, ViSA already includes sparse local selection. The distinction is the selection criterion and its role in the computation. RAR ranks patches by r_i = cos(l_i, zI) − cos(l_i, zV). The selected set determines both the local summary and the local keys prioritized during prompt refinement. The context router combines the identity anchor, selected-local summary, and stronger view-suppressed context. VSLR subsequently retains all local tokens and calibrates them through learned identity-conditioned and view-conditioned gates before bidirectional aggregation.

We assess these decisions through controlled comparisons. In Table 5, DPI, context routing, the view-affinity bias, and VSLR are fixed. Replacing identity-only dense scoring with relative dense scoring increases CARGO ALL/A↔G mAP from 65.04/69.18 to 65.57/69.87; adding Top-25% admission further increases it to 66.36/70.74. In revised Table 11, VDIA-only serves as an architectural reference. The other four variants keep the full architecture, losses, and descriptor construction fixed, varying only candidate contexts and fusion. Replacing three-context equal fusion with learned routing improves ALL/A↔G mAP by 0.58/0.58 points on CARGO and A→G/G→A mAP by 0.41/0.43 points on LAGPeR.

These comparisons isolate scoring, admission, and routing decisions within a shared architecture; they do not attribute the entire performance gap over SeCap or ViSA to a single component.

Response to Weakness 2: View affinity and useful identity evidence

We agree that view-correlated clothing and body cues can remain discriminative. At a fixed identity affinity, increasing view affinity does lower the score. This is a preference built into the selection rule, not a claim that every view-correlated cue lacks identity information. A token with high affinity to both references can still be selected if its relative score ranks sufficiently high.

This criterion is reasonable when the identity anchor provides a useful matching reference and the view representation captures variation that may distract prompt adaptation. It favors identity affinity relative to view affinity, without assuming perfect separation. Useful identity cues that are strongly view-correlated may therefore receive lower priority.

RAR restricts the local summary and prioritizes selected keys during prompt refinement, while VSLR receives the full original patch set. Calibration adds learned residual corrections to each original token. The gates depend on the token and both references, rather than view affinity alone, and are optimized through the image-level training objective. A patch excluded by RAR can therefore still contribute to local aggregation, although calibration does not guarantee that every useful cue is preserved without attenuation.

Table 5 tests the scoring preference directly: under dense admission, relative scoring improves CARGO ALL/A↔G mAP over identity-only scoring by 0.53/0.69 points; sparse admission adds another 0.79/0.87 points. Table 6 further compares separately trained selection strategies at the same token budget. Highest-score selection leads Rank-1 in all four protocols and mAP in three, with LAGPeR A→G mAP 0.05 points below random selection. The revised Sections 3.2.2 and 3.3 accordingly describe relative evidence priority and learned token calibration, without treating view affinity as a universal measure of identity irrelevance.

AuthorsOpenReview ↗

Response to Reviewer y48o

Response to Weakness 3: Local complementarity and dependence on the identity reference

We agree that retaining all tokens establishes coverage, not recovery from an inaccurate anchor. The local branch receives identity-classification and triplet supervision in Eq. 15 in addition to local-to-anchor alignment, and its calibration adds learned residual corrections to the original tokens. We have added two analyses to test complementarity and reference sensitivity.

First, Section 4.5 and Table 8 compare global, local, and fused retrieval from the same checkpoint and count both corrected and newly introduced Rank-1 errors. Results are reproduced below as Rank-1 / mAP (%); the fused column reproduces the original paper scores.

ProtocolGlobalLocalFusedGlobal wrong → fused rightGlobal right → fused wrong
AG-ReID.v2 A→C89.56 / 82.7888.41 / 81.6690.03 / 83.0318/246 (7.32%)7/2110 (0.33%)
AG-ReID.v2 C→A89.67 / 82.1488.02 / 80.1289.18 / 82.126/187 (3.21%)15/1624 (0.92%)
LAGPeR A→G40.91 / 29.8339.89 / 28.4141.33 / 29.9366/1800 (3.67%)53/1246 (4.25%)
LAGPeR G→A35.42 / 33.3133.62 / 31.7035.46 / 33.3449/1967 (2.49%)48/1079 (4.45%)

The transition columns have different denominators, so their percentages are not directly compared. Fusion produces net changes of +11, −9, +13, and +1 correct queries, respectively. Although the local descriptor is weaker on average, fusion corrects some global errors in every protocol. On AG-ReID.v2 C→A, however, fusion introduces more errors than it corrects. Global retrieval failure is a proxy for difficult matching, not a ground-truth label of anchor error.

Second, Appendix A.3.4 and Table 13 test sensitivity to the shared identity reference at fixed checkpoints. We interpolate it with a fixed different-identity reference from the same coarse view, preserving its original norm. The same perturbed reference is used by DPI, RAR, context routing, and VSLR; model parameters, patch tokens, view features, the final global descriptor, and gallery descriptors remain fixed. Across the five intervention strengths, local and fused mAP change by at most 0.02 and 0.01 percentage points at the reported precision. The local results show that this stability is not solely due to the unchanged global component. The original and fully replaced references remain geometrically close on AG-ReID.v2 (mean cosine about 0.992), limiting the strength of that test; the corresponding value is about 0.4 on LAGPeR.

These results show complementary matching information on some queries and low sensitivity to the tested reference replacements. They do not establish detection or repair of incorrect anchors, or isolate the training-time effect of alignment.

Thank you again for your time and constructive feedback. We would be happy to clarify any remaining points or answer further questions.

Thank you to the author for responding to my questions. The author's response has addressed most of my concerns. I will keep my score unchanged.

AuthorsOpenReview ↗

Response to Reviewer y48o

We sincerely thank the reviewer for carefully considering our response. We are pleased that our clarifications have addressed most of your concerns. We greatly appreciate your thoughtful assessment and valuable feedback, which will help us further improve our work.

Reviewer_4o2JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Reliability-Adaptive Prompt Routing (RAPR) for aerial-ground person re-identification. The method constructs a view-decoupled identity anchor and integrates dynamic prompt initialization, identity-versus-view similarity-based local token selection, and view-suppressed local feature refinement to improve cross-view identity discrimination. Experiments on CARGO, AG-ReID.v2, and LAGPeR demonstrate the effectiveness of the proposed framework.

Critical strengths
  1. The paper addresses a practically relevant problem. It investigates viewpoint variations and local feature reliability in aerial-ground person re-identification, which is an important and challenging research problem.
  2. The proposed framework is well structured. It integrates dynamic prompt learning, reliability-guided local feature selection, and view-suppression mechanisms, with experiments and ablation studies conducted on multiple datasets to evaluate the contributions of its components.
Critical weaknesses
  1. This paper uses the similarity gap between identity anchors and viewpoint features to determine which local evidence can be incorporated into prompt updates. What innovations does this approach offer compared to existing viewpoint decoupling, component attention, and adaptive prompt methods? Furthermore, from Figure 1, it appears that only viewpoint differences are being captured. Could component attention methods also address this issue? 2.Most RAPR methods are designed with a specific viewpoint in mind. There are no relevant methods designed to address issues such as changes in scale, visibility, and background. The methods are not well-suited to the problems they are intended to solve. 3.The paper uses the similarity gap between local tokens and identity anchors, as well as viewpoint features, as a reliability metric; however, it has not yet been fully demonstrated that this metric has a stable correlation with the actual identity discrimination capability of the local region.
Questions for authors

See Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer 4o2J

Thank you for your careful review. We appreciate your recognition of the practical importance of the problem, the structure of our framework, and the experiments across multiple datasets. We have clarified the methodological contribution and added evidence to assess the reliability criterion.

Response to Weakness 1: Novelty and distinction from part-based attention

RAPR builds on view-decoupled encoders and prompt learning. SeCap recalibrates prompts using the view-decoupled global representation and performs bidirectional prompt–local refinement. ViSA generates semantic tokens through shared and view-specific experts, then selects Top-k local tokens by cosine similarity to the expert queries for graph-based fusion.

RAPR's contribution is to use paired identity and view references for two linked decisions: which local evidence updates the prompt, and how all local tokens are calibrated before aggregation. RAR ranks patches by their relative affinity to the two references; the selected set determines the local summary and the local keys prioritized during prompt refinement. The router combines the identity anchor, selected-local summary, and stronger view-suppressed context. VSLR then retains all local tokens and calibrates them through learned identity-conditioned and view-conditioned gates.

Part-based attention can also help address the difficulties in Figure 1. A part partition or foreground mask alone, however, does not specify how to balance a token's identity and view affinities. RAPR makes this comparison before prompt updating, without requiring fixed body-part correspondence.

Table 5 tests scoring and admission with DPI, context routing, the view-affinity bias, and VSLR held fixed. On CARGO ALL/A↔G, replacing identity-only dense scoring with relative scoring improves mAP by 0.53/0.69 points; Top-25% admission adds another 0.79/0.87 points. The four controlled rows below the VDIA-only reference in revised Table 11 isolate context composition and fusion within the full architecture: learned routing improves CARGO ALL/A↔G mAP by 0.58/0.58 points over three-context equal fusion.

Response to Weakness 2: Relation to scale, visibility, and background changes

RAPR uses sample-dependent feature selection and calibration to address the potential effects of scale, visibility, and background changes on the available identity evidence. It does not explicitly estimate object scale, body-part visibility, or foreground masks, and we do not claim invariance to each factor.

  • Scale and resolution: changes in scale may reduce fine garment details while preserving coarser color and silhouette cues. RAR combines a global identity anchor with a score-selected local summary and predicts their routing weights for each image. This provides an indirect way to adjust the balance between global and local evidence.
  • Visibility and occlusion: different images expose different body regions. RAR selects from the observed tokens without requiring fixed body-part correspondence. VSLR then calibrates the full patch set, so exclusion from the sparse prompt-update set does not remove a region from subsequent local aggregation.
  • Background and view-dependent clutter: RAR assesses each token against both identity and view references before prompt refinement. VSLR applies learned identity-conditioned and view-conditioned gates to calibrate local memory before aggregation.

The revised introduction explains why these imaging changes motivate selecting and calibrating local evidence; Sections 3.2–3.3 describe the corresponding computations. Our benchmark results assess overall retrieval performance and do not independently quantify robustness to scale, visibility, or background changes.

AuthorsOpenReview ↗

Response to Reviewer 4o2J

Response to Weakness 3: Reliability scoring and evidence selection

We have clarified the scoring rule and added two experiments to Section 4.3: a fixed-budget selection comparison (Table 6) and a fixed-checkpoint evidence-replacement analysis (Figure 3).

The score r_i = cos(l_i, zI) − cos(l_i, zV) ranks each token by its affinity to the identity anchor relative to the view reference. RAR uses this difference to select and weight local evidence before it updates the prompt.

First, Table 5 separates the effects of relative scoring and sparse admission, with DPI, context routing, the view-affinity bias, and VSLR held fixed. Under dense admission, replacing identity-only scoring with relative scoring increases CARGO ALL/A↔G mAP from 65.04/69.18 to 65.57/69.87, gains of 0.53/0.69 points. Applying Top-25% admission further increases mAP to 66.36/70.74, adding 0.79/0.87 points. A↔G Rank-1 remains 69.15% in the first comparison and rises to 70.21% with sparse admission.

Second, Table 6 compares highest-r, random, and lowest-r selection, each retaining 25% of the tokens. The variants are trained separately, varying only the token-selection strategy. All entries below are Rank-1 / mAP (%).

Selection strategy (25%)CARGO ALLCARGO A↔GLAGPeR A→GLAGPeR G→A
Highest r (original RAPR)73.08 / 66.3670.21 / 70.7441.33 / 29.9335.46 / 33.34
Random68.27 / 59.7865.96 / 65.8441.07 / 29.9834.57 / 33.21
Lowest r67.95 / 60.8669.15 / 69.2240.71 / 29.9234.31 / 33.03

Highest-r selection achieves the best Rank-1 in all four protocols and the best mAP in three. On CARGO ALL/A↔G, it improves mAP over random selection by 6.58/4.90 points and over lowest-r selection by 5.50/1.52 points. On LAGPeR, Rank-1 improves in both directions, while A→G mAP is 0.05 points below random selection.

Third, Figure 3 examines the original admitted set at a fixed checkpoint. Equal-size lower, central, or upper score windows are replaced by shared, previously unselected tokens. The new weights update both the summary and prompt interaction, while VSLR retains all tokens. Recomputed softmax and uniform weights each use their own unreplaced query baseline; gallery descriptors remain fixed. Upper-window replacement produces larger mean AP losses on CARGO and LAGPeR A→G, but a mean gain on LAGPeR G→A. All upper-window intervals include zero.

Tables 5 and 6 support the score for selecting evidence for prompt updates; Figure 3 tests sensitivity within the trained model, without establishing a universal ordering of individual token contributions across protocols. We have clarified this scope in Section 3.2.2 and aligned the introduction, overview, Appendix B.2, and conclusion accordingly: a higher score gives a token greater priority for prompt updating, without guaranteeing a larger individual contribution to identity discrimination. The scoring formula and model computations are unchanged.

Thank you again for your careful review and helpful comments. Please let us know if you have any further questions or would like additional clarification.

AuthorsOpenReview ↗

Response to Reviewer 4o2J

Thank you again for your careful review. With less than a day left in the discussion period, we wanted to briefly follow up on our responses to your review. We’re sorry to bother you so close to the deadline.

In our responses, we addressed how RAPR differs from prior methods, how changes in scale, visibility, and background relate to our approach, and what the added experiments show about using the reliability score to select local evidence.

If you have a chance to look over our responses, we’d be grateful for any feedback. We’d be happy to clarify any points that remain open. Thank you for your time.

Dear Reviewer,

Please respond to the authors’ rebuttal on OpenReview. Briefly indicate whether the rebuttal addresses your concerns and confirm or revise your score accordingly.

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.