acceptodds
Under review as a conference paper at ICLR 2027

Bounded Contextual Refinement with Neighborhood Support for Text-Aerial Person Retrieval

Abstract

Text-aerial person retrieval aims to retrieve images of a target person from an aerial gallery given a natural language description. Variations in viewing angle, altitude, and occlusion can obscure the appearance cues described in the query, making visually similar individuals difficult to distinguish. Existing methods improve image-text alignment, but scoring each gallery image independently does not fully exploit context from other gallery images. In this paper, we propose Bounded Contextual Refinement (BCR), which improves aerial representations and retrieval rankings through bounded feature calibration and reciprocal gallery context. Bounded Feature Calibration (BFC) learns a norm-constrained residual transformation by contrasting the highest-scoring positive and negative gallery images for each text query. The constraint limits the magnitude of each feature correction while allowing the representation to adapt. Reciprocal Contextual Ranking (RCR) addresses two distinct retrieval decisions. It constructs candidate context vectors from reciprocal neighborhoods and learns bounded score corrections for candidate selection among the initial top five candidates. It then combines matching scores with neighborhood support to rank the remaining candidates within fixed intervals. To strengthen supervision for aerial appearance, Auxiliary Description Supervision (ADS) supplements training of the CLIP encoder with descriptions generated from aerial training images and reviewed against them by the same frozen vision-language model. Experiments on AERI-PEDES and TBAPR show improvements in Rank-1 accuracy and mAP over the CFAN baseline. The code will be made publicly available.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Text-based aerial person retrieval, Contextual re-ranking, Shortlist preservation, Bounded feature calibration, Reciprocal gallery context

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 3 reviews, above 85% of the others.
Reviewer_ZTF3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Bounded Contextual Refinement (BCR) for text-aerial person retrieval. It combines bounded residual feature calibration, reciprocal gallery context for candidate selection and ranking, and auxiliary supervision from generated aerial descriptions. The framework addresses appearance ambiguity in aerial imagery and demonstrates improvements on AERI-PEDES and TBAPR.

Critical strengths
  1. The method effectively connects incomplete aerial appearance cues with the need for contextual evidence from other gallery images. Separating top-candidate selection from subsequent ranking is well motivated.
  2. Bounding feature updates and preserving the initial top-five and top-ten candidate sets provide explicit control over the refinement process.
  3. On AERI-PEDES, BCR improves Rank-1 from 47.16% to 49.29% and mAP from 44.79% to 47.70% over CFAN with ground-view supervision. The improvements in mAP and mINP on both benchmarks further support the value of contextual ranking.
  4. The paper proposes a targeted solution to key challenges in text-to-aerial person retrieval, demonstrating a degree of novelty.
Critical weaknesses
  1. The paper specifies the residual bound, reciprocal-neighborhood size, and neighborhood mixing coefficient, but barely explains their selection. Please describe the selection criteria and validation procedure, and whether both datasets share these settings.
  2. Section 3.4 states that the same frozen vision-language model generates descriptions and checks clothing, footwear, and carried objects against the image, removing unsupported attributes. Please provide an example with the source image, initial and reviewed descriptions, highlighting removed or corrected attributes and explaining how ambiguous, occluded, or unverifiable details are handled.
Questions for authors
  1. The paper specifies the residual bound, reciprocal-neighborhood size, and neighborhood mixing coefficient, but barely explains their selection. Please describe the selection criteria and validation procedure, and whether both datasets share these settings.
  2. Section 3.4 states that the same frozen vision-language model generates descriptions and checks clothing, footwear, and carried objects against the image, removing unsupported attributes. Please provide an example with the source image, initial and reviewed descriptions, highlighting removed or corrected attributes and explaining how ambiguous, occluded, or unverifiable details are handled.
  3. Figure 2 combines description generation, encoder adaptation, feature calibration, and contextual scoring, but some text and arrow labels are small. Please enlarge key annotations and indicate the training order, frozen modules at each stage, and inference pathway.
  4. The method section would benefit from clearer exposition. The numerous equations and limited intuitive explanations make the approach difficult to follow. I suggest highlighting the core ideas and adding explanatory text where needed.
5 comments
AuthorsOpenReview ↗

Response to Reviewer ZTF3

W1: The paper specifies the residual bound, reciprocal-neighborhood size, and neighborhood mixing coefficient, but barely explains their selection. Please describe the selection criteria and validation procedure, and whether both datasets share these settings.

A1: Thank you for the question. Both datasets use ρ = 0.1, k = 4, α = 0.25, M = 5, L = 64, c₁ = 0.05, c₂ = 0.02 and intervals [2, 5], [6, 10], [11, N]. Each model is trained on its own dataset's released split and CFAN initialization; the refinement settings are shared.

These values limit the feature update, keep neighborhood support local and give each image's own score the larger weight in interval ordering. The selection depth and intervals preserve the top-five and top-ten sets. Tables C, G and H test sensitivity around these fixed defaults.

For validation, we excluded one of three training-identity folds—1,174 identities on AERI-PEDES and 393 on TBAPR—from both scorer-fitting queries and candidate galleries, then evaluated on that fold with its own gallery. The encoder and BFC remain trained on the full training split, so this validates the scoring stage conditional on those representations. No test identities are used.

The held-out identities are excluded from both the queries and candidate galleries used to fit the scorers. This directly tests refinement on identities outside scorer fitting while retaining the learned representation. The sensitivity tables vary one parameter at a time around the same defaults; the validation table checks whether the default refinement and nearby k/α values also improve ranking on this separate training-identity fold.

Table J: Scorer validation on training identities excluded from scorer fitting (%; encoder and BFC fixed).

SettingAERI-PEDES Rank-1AERI-PEDES mAPTBAPR Rank-1TBAPR mAP
Direct ranking69.3667.9563.0855.57
Default refinement69.8768.8863.9556.19
k = 369.8668.7264.0856.07
k = 569.9268.9564.0856.30
α = 0.2069.8768.7463.9556.09
α = 0.3069.8769.0063.9556.28

The defaults improve validation Rank-1/mAP by +0.51%/+0.93% on AERI-PEDES and +0.87%/+0.61% on TBAPR, with similar gains at nearby k and α values. Section 4.2 specifies the shared settings, and Appendix I describes the validation procedure.

W2: Section 3.4 states that the same frozen vision-language model generates descriptions and checks clothing, footwear, and carried objects against the image, removing unsupported attributes. Please provide an example with the source image, initial and reviewed descriptions, highlighting removed or corrected attributes and explaining how ambiguous, occluded, or unverifiable details are handled.

A2: Thank you for this suggestion. The same frozen Qwen3-VL-8B-Instruct model first drafts a description and then checks it against the same aerial training image. Neither pass receives ground-view images or released annotations. The review prompt asks the model to retain coarser visible categories when fine details are unclear and to omit occluded or unverifiable attributes.

We reproduced the following pair on an AERI-PEDES training image using the retained model, prompts and deterministic decoding. The reviewed text matches the cached description used for training.

Draft: "Person wears a dark short-sleeved shirt, light blue jeans, and dark sneakers with white soles. Carries a brown bag with a visible pattern over their right shoulder."

Reviewed: "Person wears a dark short-sleeved shirt, light blue jeans, and dark footwear. Carries a brown bag with a discernible pattern over their right shoulder."

The review replaces “dark sneakers with white soles” with “dark footwear” because the shoe type and sole color cannot be verified at this resolution. It retains the visible coarse category. The change from “visible pattern” to “discernible pattern” only adjusts the wording.

The review prompt ties fine-grained attributes to visible image evidence and requests a coarser category or omission when that evidence is absent. In this example, the useful evidence that the person wears dark footwear is preserved while unsupported shoe-type and sole-color detail is removed. This keeps the auxiliary description focused on the appearance information available in the aerial view.

Figure A9 in Appendix F.4 shows the source image and both descriptions, with removed details in red and the retained category in blue. Section 3.4 explains how uncertain attributes are handled. Reviewed descriptions provide auxiliary supervision; the released annotations remain the primary training data.

AuthorsOpenReview ↗

Response to Reviewer ZTF3

Q1: The paper specifies the residual bound, reciprocal-neighborhood size, and neighborhood mixing coefficient, but barely explains their selection. Please describe the selection criteria and validation procedure, and whether both datasets share these settings.

A1: Thank you. This question corresponds to W1, and the added validation results are reproduced in Table J below. Both datasets use ρ = 0.1, k = 4, α = 0.25, M = 5, L = 64, c₁ = 0.05 and c₂ = 0.02, with intervals [2, 5], [6, 10], [11, N]. The design limits feature and score changes, favors local support and the image’s own score, and preserves the initial top-five and top-ten sets.

For validation, one training-identity fold is excluded from both scorer-fitting queries and candidate galleries: 1,174 identities on AERI-PEDES and 393 on TBAPR. The encoder and BFC remain fixed after training on the full training split. This is conditional validation of the scoring stage, and no test identities are used.

Table J: Scorer validation on training identities excluded from scorer fitting (%; encoder and BFC fixed).

SettingAERI-PEDES Rank-1AERI-PEDES mAPTBAPR Rank-1TBAPR mAP
Direct ranking69.3667.9563.0855.57
Default refinement69.8768.8863.9556.19
k = 369.8668.7264.0856.07
k = 569.9268.9564.0856.30
α = 0.2069.8768.7463.9556.09
α = 0.3069.8769.0063.9556.28

Default refinement improves validation Rank-1/mAP by +0.51%/+0.93% on AERI-PEDES and +0.87%/+0.61% on TBAPR, using unrounded metrics. Nearby k and α values retain similar gains. Section 4.2 specifies the shared settings, Appendix H reports all sensitivity tests, and Appendix I describes the validation protocol and results.

Q2: Section 3.4 states that the same frozen vision-language model generates descriptions and checks clothing, footwear, and carried objects against the image, removing unsupported attributes. Please provide an example with the source image, initial and reviewed descriptions, highlighting removed or corrected attributes and explaining how ambiguous, occluded, or unverifiable details are handled.

A2: Thank you. The example introduced in W2 is reproduced here. The same frozen Qwen3-VL-8B-Instruct model first drafts a description and then reviews it against the same aerial training image. Neither pass receives a ground-view image or a released annotation. The retained model and prompts, with deterministic decoding, reproduce the reviewed description stored in the training cache.

Draft: "Person wears a dark short-sleeved shirt, light blue jeans, and dark sneakers with white soles. Carries a brown bag with a visible pattern over their right shoulder."

Reviewed: "Person wears a dark short-sleeved shirt, light blue jeans, and dark footwear. Carries a brown bag with a discernible pattern over their right shoulder."

The review retains the visible coarse category “dark footwear” and removes the unverifiable shoe type and sole color. The prompt asks for coarser visible categories for ambiguous details and omission of occluded or unverifiable attributes. The change from “visible pattern” to “discernible pattern” is wording only. Figure A9 in Appendix F.4 pairs the source image with both descriptions and highlights removed and retained attributes. Section 3.4 explains the procedure. The reviewed text provides auxiliary supervision alongside the released primary annotations.

AuthorsOpenReview ↗

Response to Reviewer ZTF3

Q3: Figure 2 combines description generation, encoder adaptation, feature calibration, and contextual scoring, but some text and arrow labels are small. Please enlarge key annotations and indicate the training order, frozen modules at each stage, and inference pathway.

A3: Thank you for the helpful suggestions. Figure 2 identifies the four training stages. Sections 3.1–3.4 clarify the training order, the trainable and frozen modules, and the inference procedure. Table K summarizes these stages.

Table K: Training stages and inference of BCR.

StageTrained or generatedFrozen or reused
1. Prepare ADS descriptionsDraft and reviewed descriptions, generated once and cachedVision-language model
2. Adapt the encoderCLIP encoder initialized from CFAN, with primary and auxiliary objectives (5 epochs)Cached descriptions
3. Fit BFCResidual matrix W (49 epochs)Encoder
4. Fit RCRh₁ (20 epochs), then h₂ (20 epochs)Encoder and BFC; also h₁ while fitting h₂
InferenceNo training; text encoding, scoring against the cached calibrated gallery, candidate selection and interval orderingAll modules; no description generation

The freezing order makes the dependency between stages explicit. BFC is fitted on the adapted encoder’s fixed features. Both RCR scorers then use fixed encoder/BFC outputs, and h₁ is additionally frozen before h₂ is fitted. At inference, the gallery encoding, calibration and gallery graph can be reused across queries; the vision-language model is absent from the retrieval path. These distinctions separate offline preparation, parameter learning and query-time computation.

Inference uses the adapted text encoder, scores the cached calibrated gallery, selects the first result and orders the remaining candidates within intervals. Section 3.1 clarifies that inference processes every query without description generation or identity labels; training-query filtering is used only to fit the scorers.

Q4: The method section would benefit from clearer exposition. The numerous equations and limited intuitive explanations make the approach difficult to follow. I suggest highlighting the core ideas and adding explanatory text where needed.

A4: Thank you for this suggestion. BCR first makes a small adjustment to aerial features to improve identity discrimination, then uses neighboring gallery images to support the ranking. This provides a simple intuition for BFC and RCR before introducing their equations.

In BFC, the residual determines the update direction, the norm cap limits its size and the identity connection retains the input feature. In RCR, a candidate receives stronger support when visually related images also match the query. Bounded corrections select the first result, and neighborhood evidence orders the remainder within intervals that preserve the initial top-five and top-ten sets.

The two scorers also have an intuitive division of roles: h₁ reads a candidate’s context directly, while h₂ measures it relative to the other selection candidates through centered context features. The first-result decision and interval ordering therefore use context for different purposes: choosing the leading match and organizing the remaining relevant images. This explanation accompanies the mathematical definitions in Section 3.3.

We added these explanations before and alongside the equations in Sections 3.2 and 3.3.

Reviewer_668eOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed Bounded Contextual Refinement (BCR) for text-aerial person retrieval, which aims to retrieve aerial images of a target person given a natural language description. The authors address the challenges of viewpoint variations, occlusions, and visually similar individuals by incorporating gallery-level contextual information into conventional image-text matching. The proposed framework consists of three components: (1) Auxiliary Description Supervision (ADS), which uses a frozen vision-language model to generate and review aerial-image descriptions for additional CLIP encoder supervision; (2) Bounded Feature Calibration (BFC), which learns norm-constrained residual transformations to improve the discrimination between positive and negative candidates while limiting feature perturbation; and (3) Reciprocal Contextual Ranking (RCR), which leverages reciprocal gallery neighborhoods to refine retrieval results. RCR separately performs top-ranked candidate selection using learned bounded score corrections and interval-constrained re-ranking using neighborhood support, preserving the initial top-5 and top-10 candidate sets. The authors evaluate BCR on two text-aerial person retrieval benchmarks, AERI-PEDES and TBAPR, reporting improvements over the CFAN baseline in Rank-1 accuracy and mean Average Precision (mAP). On AERI-PEDES, BCR achieves 49.29% Rank-1 accuracy and 47.70% mAP, compared with 47.16% and 44.79% for CFAN with ground-view supervision. On TBAPR, BCR achieves 49.72% Rank-1 accuracy and 44.87% mAP. The paper further provides component-wise ablation studies, repeated training experiments, theoretical analyses of bounded feature perturbations and candidate-set preservation, and computational and qualitative evaluations to examine the effectiveness and limitations of the proposed framework.

Critical strengths

S1. The paper addresses a well-motivated limitation of existing text-aerial person retrieval methods by incorporating gallery-level contextual information into the retrieval process. The proposed approach is particularly relevant to aerial imagery, where viewpoint variations and occlusions make individual image-text matching scores unreliable.

S2. The bounded feature calibration module provides a principled mechanism for adapting pretrained representations while controlling the magnitude of feature perturbations. The theoretical bound is straightforward but valid, and the ablation results support the benefit of norm-constrained residual adaptation over its unbounded counterpart.

Critical weaknesses

W1. The methodological novelty appears somewhat limited, as the proposed framework combines established ideas, including residual feature adaptation, auxiliary caption supervision, and neighborhood-based re-ranking. Although their integration is reasonable, the paper does not sufficiently distinguish the proposed contextual refinement mechanism from conventional k-reciprocal re-ranking or other neighborhood-aware correction methods under controlled experimental conditions.

W2. The effectiveness of several proposed components is not consistently supported by the repeated experiments. In particular, BFC slightly decreases mAP compared with omitting calibration, and the second contextual scorer consistently degrades both Rank-1 and mAP across all three reported seeds. These results raise concerns about whether the full architecture is justified by its empirical contributions.

W3. Several important design choices, including the reciprocal neighborhood size, candidate selection depth, contextual correction bounds, and fixed ranking intervals, appear heuristic. The paper would benefit from a systematic sensitivity analysis to establish whether the reported improvements are robust to these choices.

Questions for authors

Q1. The repeated experiments in Appendix Table A5 indicate that adding the second contextual scorer h2h_2 consistently decreases both Rank-1 accuracy and mAP across all three seeds, despite showing improvements in the main checkpoint. Could the authors explain why h2h_2 is retained in the final framework and provide further evidence supporting its contribution? In particular, what additional information does the centered contextual representation capture that cannot be effectively modeled by the first scorer h1h_1?

Q2. A central motivation of the proposed RCR module is that reciprocal neighborhoods provide more reliable contextual information than individual gallery similarities. However, the paper does not directly compare reciprocal neighborhoods with standard k-nearest-neighbor aggregation or similarity-weighted neighborhood support. Could the authors clarify whether the observed improvements primarily arise from the reciprocal neighborhood constraint or from incorporating neighborhood information in general? A controlled comparison using identical features and ranking procedures would help establish the necessity of reciprocity.

Q3. The proposed framework shares conceptual similarities with existing neighborhood-aware correction and k-reciprocal re-ranking approaches. Could the authors clarify the key methodological differences between RCR and these existing methods? In particular, would conventional neighborhood-based re-ranking applied to the same CFAN representations achieve comparable improvements? Such a comparison would help assess the specific contribution of the proposed contextual refinement mechanism.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 668e

W1: The methodological novelty appears somewhat limited, as the proposed framework combines established ideas, including residual feature adaptation, auxiliary caption supervision, and neighborhood-based re-ranking. Although their integration is reasonable, the paper does not sufficiently distinguish the proposed contextual refinement mechanism from conventional k-reciprocal re-ranking or other neighborhood-aware correction methods under controlled experimental conditions.

A1: Thank you for the question. RCR separates first-result selection from the ordering of remaining candidates. Two learned corrections, bounded by 0.05 and 0.02, select the first result from the initial top five; neighborhood support then orders the remainder within fixed intervals. This preserves the initial top-five and top-ten sets (Proposition B.2). Conventional k-reciprocal re-ranking applies a revised distance to the entire list.

RCR uses a gallery-only graph built once for all queries and fits 28 scorer weights on training identities. The conventional method is nonparametric and builds a joint query–gallery graph. These differences allow RCR to learn bounded corrections while processing each query independently.

We compared RCR with the authors' k-reciprocal implementation (Zhong et al., CVPR 2017), using its published defaults (k₁ = 20, k₂ = 6, λ = 0.3) on the same features and initial similarities. Table E includes both CFAN and calibrated features. No method uses test identity labels.

Table E: Re-ranking comparison on identical features (%).

DatasetFixed featuresRankingRank-1Rank-5Rank-10mAP
AERI-PEDESCFANDirect47.4565.7273.7845.44
AERI-PEDESCFANk-reciprocal45.7464.6873.6745.54
AERI-PEDESADS + BFCDirect48.6766.1174.1746.10
AERI-PEDESADS + BFCk-reciprocal45.7664.2173.6246.07
AERI-PEDESADS + BFCRCR (ours)49.2966.1174.1747.70
TBAPRCFANDirect49.3866.0073.0044.02
TBAPRCFANk-reciprocal44.6362.9171.2543.04
TBAPRADS + BFCDirect49.5666.7572.7843.79
TBAPRADS + BFCk-reciprocal45.3162.9170.8442.87
TBAPRADS + BFCRCR (ours)49.7266.7572.7844.87

The CFAN rows evaluate conventional re-ranking on the initial representation. The ADS + BFC rows compare direct ranking, conventional re-ranking and RCR on one fixed calibrated representation. Thus the differences within those rows come from the ranking procedure. This complements the structural distinction: RCR learns a bounded first-result decision while retaining the initial candidate sets, and uses neighborhood evidence to improve ordering within them.

On calibrated features, RCR gives Rank-1/mAP gains of +0.62%/+1.60% on AERI-PEDES and +0.16%/+1.08% on TBAPR. Under its published defaults, k-reciprocal re-ranking reduces Rank-1 and does not improve mAP on these features. It also changes the top-five set for 32% and 29% of queries, respectively. RCR improves both metrics while preserving these sets. We clarified this distinction in Section 3.3 and added the comparison to Appendix G.2.

AuthorsOpenReview ↗

Response to Reviewer 668e

W2: The effectiveness of several proposed components is not consistently supported by the repeated experiments. In particular, BFC slightly decreases mAP compared with omitting calibration, and the second contextual scorer consistently degrades both Rank-1 and mAP across all three reported seeds. These results raise concerns about whether the full architecture is justified by its empirical contributions.

A2: Thank you for examining these results. Across the three training repeats, bounded BFC outperforms its unbounded counterpart on all five metrics in every run, and complete RCR improves both Rank-1 and mAP over direct ranking. Each run repeats encoder adaptation, BFC and RCR from the same CFAN initialization. Table F reports the paired differences.

Table F: Paired metric differences (Δ, %) over three AERI-PEDES training repeats, reported as mean ± sample standard deviation.

ComparisonΔ Rank-1Δ mAP
Bounded vs. unbounded BFC+0.489 ± 0.169+0.612 ± 0.048
Complete RCR vs. direct ranking+0.494 ± 0.179+1.578 ± 0.037
BFC vs. no BFC+0.282 ± 0.307-0.073 ± 0.040
h₁ + h₂ vs. h₁ (intervals retained)-0.065 ± 0.102+0.005 ± 0.013

For each seed, we compute the difference between paired variants before averaging across seeds; the ± values are sample standard deviations of those paired differences. The complete training repeats and the fixed-representation scorer refits answer complementary questions. The former evaluate the components through the full training pipeline, while the latter isolate the contribution of scorer training on a retained representation.

BFC optimizes the separation of the highest-scoring positive and negative candidates. Its +0.68% Rank-1 gain before RCR and the paired Rank-1 gains in Table F support this targeted role. Rank-1 measures the first result, whereas mAP also depends on the positions of all relevant gallery images. The norm cap bounds feature displacement by ‖b − a‖₂ ≤ 2ρ and improves mean mAP by +0.612% over unbounded calibration. Both residual variants use the same parameterization, so this comparison isolates the contribution of the norm constraint.

The second scorer provides a smaller relative correction. Its paired contribution is reported in Table F; the retained AERI-PEDES model shows gains in both metrics (Table A), and four scorer refits consistently improve AERI-PEDES mAP (Table I). Our response to Q1 below explains its role and this additional evidence.

AuthorsOpenReview ↗

Response to Reviewer 668e

W3: Several important design choices, including the reciprocal neighborhood size, candidate selection depth, contextual correction bounds, and fixed ranking intervals, appear heuristic. The paper would benefit from a systematic sensitivity analysis to establish whether the reported improvements are robust to these choices.

A3: Following your suggestion, we tested each refinement parameter. Table C covers k and α; Table G covers selection depth M, context depth L, correction caps c₁, c₂ and residual cap ρ. Each test changes one parameter while keeping the fitted weights fixed, so it measures sensitivity of the existing model.

Table G: Parameter sensitivity (%). Defaults are bold; one parameter is changed at a time.

ParameterValueAERI-PEDES Rank-1AERI-PEDES mAPTBAPR Rank-1TBAPR mAP
Selection depth M349.2147.6749.7844.86
Selection depth M449.2847.6949.7244.86
Selection depth M549.2947.7049.7244.87
Selection depth M749.2647.6949.7544.87
Selection depth M1049.1947.6749.7844.87
Context depth L3249.3147.6849.6644.86
Context depth L4849.2847.6849.7244.87
Context depth L6449.2947.7049.7244.87
Context depth L9649.1947.6849.7844.88
Context depth L12849.1847.6949.7844.88
Correction cap c₁0.02549.2947.6749.8144.89
Correction cap c₁0.037549.3947.6949.7544.87
Correction cap c₁0.0549.2947.7049.7244.87
Correction cap c₁0.062549.1147.6749.7544.86
Correction cap c₁0.07549.1347.6949.8144.88
Correction cap c₂049.2147.6750.0044.88
Correction cap c₂0.0149.3447.6949.7844.87
Correction cap c₂0.01549.3647.6949.6944.87
Correction cap c₂0.0249.2947.7049.7244.87
Correction cap c₂0.02549.1947.6949.8444.88
Correction cap c₂0.0349.1347.6849.8844.90
Residual cap ρ0.0549.2947.8850.1645.12
Residual cap ρ0.07549.2647.7949.9145.06
Residual cap ρ0.149.2947.7049.7244.87
Residual cap ρ0.12549.2447.6649.7544.79
Residual cap ρ0.1549.2447.6649.7844.79
Residual cap ρ0.249.2447.6649.7844.79

Every setting in Table G exceeds direct ranking on both Rank-1 and mAP. The total Rank-1/mAP ranges are 0.28%/0.22% on AERI-PEDES and 0.50%/0.33% on TBAPR, showing limited variation around the defaults.

These parameters control distinct parts of refinement: M sets the first-result candidate pool, L sets the context depth, ρ bounds feature calibration, and c₁/c₂ bound the two score corrections. All sweeps use the retained weights. Increasing M to 7 or 10 changes the initial top-five set for 0.98–1.11% of AERI-PEDES queries and 1.38–1.75% of TBAPR queries. M = 5 therefore matches the intended candidate-set guarantee as well as the reported retrieval gains.

M = 5 is the largest selection depth that preserves the initial top-five set. Boundaries at ranks 5 and 10 similarly preserve both reported sets. In Table H, this partition gives the highest mAP among the tested set-preserving partitions. Removing a boundary can raise mAP but reduces Rank-5 or Rank-10. These results support the chosen constraints. Appendix H reports these sensitivity analyses.

Table H: Interval-boundary comparison (%). The first result is fixed across rows.

DatasetInterval boundariesRank-1Rank-5Rank-10mAP
AERI-PEDES5, 10 (default)49.2966.1174.1747.70
AERI-PEDES3, 5, 1049.2966.1174.1747.44
AERI-PEDES5, 10, 2049.2966.1174.1747.59
AERI-PEDES5, 2049.2966.1173.2047.84
AERI-PEDES1049.2964.6374.1748.10
AERI-PEDESNone (unrestricted tail)49.2964.6073.1648.38
TBAPR5, 10 (default)49.7266.7572.7844.87
TBAPR3, 5, 1049.7266.7572.7844.76
TBAPR5, 10, 2049.7266.7572.7844.80
TBAPR5, 2049.7266.7571.8144.98
TBAPR1049.7265.1672.7845.06
TBAPRNone (unrestricted tail)49.7265.1671.7845.24
Reviewer_TTQWOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Bounded Contextual Refinement (BCR) for text-aerial person retrieval. The task is to find a target person in aerial images using a text description. Changes in viewing angle, altitude, and occlusion can hide details and make different people look similar. BCR builds on CFAN and includes three parts. Auxiliary Description Supervision uses a frozen vision-language model to generate descriptions from aerial training images. The same model checks these descriptions against the images. Bounded Feature Calibration learns small changes to image features. It aims to give correct matches higher scores than incorrect matches. Reciprocal Contextual Ranking uses information from similar gallery images. It first selects the top result from the initial top five candidates using bounded score corrections. It then sorts the remaining images within fixed rank intervals. The top-five and top-ten candidate sets remain unchanged. Experiments on AERI-PEDES and TBAPR show higher Rank-1 accuracy and mAP than CFAN.

Critical strengths
  • The paper uses information from similar gallery images to support matching. Since aerial images often hide details mentioned in the text, this is a reasonable way to address the problem.

  • BCR improves Rank-1 and mAP over CFAN on both datasets. The paper includes ablation studies, repeated training runs, and failure cases. These analyses help explain the role of each component and show when the method helps or fails.

  • The language model is only used to prepare training descriptions. Comapred to the CFAN, BFC and the ranking scorers add few parameters. The reported query time increases only slightly compared with CFAN.

Critical weaknesses
  • The comparison in Table 3 needs to be clearer. Table 3 reports 48.85% Rank-1 with (h_1) and 49.29% with (h_1+h_2). These results use different settings. In Section 4.4, the results under the same controlled setting are 48.85% and 49.00%. The table therefore does not clearly show the gain from adding (h_2) alone. Please explain the differences between the main and controlled settings.

  • The benefit of reciprocal neighbors needs a direct comparison. RCR only keeps neighbors that select each other. The main paper does not compare this choice with ordinary nearest neighbors. A comparison using the same image features would help show whether reciprocal neighbors provide better support for ranking.

  • The parameter choices need more support. The main paper does not clearly explain how the neighborhood size and the weight of neighborhood support are chosen. These choices may affect the final ranking. A small test with a few nearby values would help show whether the improvements hold across different settings.

  • The effect of extra training is not clear. BCR starts from a CFAN checkpoint and trains the encoder for five more epochs. Table 2 does not report CFAN with the same extra training but without ADS, BFC, or RCR. Please add this baseline using the original CFAN training loss. This would help show how much of the improvement comes from extra training and how much comes from the proposed components.

Questions for authors

Please address the weaknesses listed above. I believe these issues can all be resolved through further explanation and additional experiments. I am generally in favor of accepting the paper and hope the authors can address these concerns in their response.

5 comments
AuthorsOpenReview ↗

Response to Reviewer TTQW

W1: The comparison in Table 3 needs to be clearer. Table 3 reports 48.85% Rank-1 with (h₁) and 49.29% with (h₁ + h₂). These results use different settings. In Section 4.4, the results under the same controlled setting are 48.85% and 49.00%. The table therefore does not clearly show the gain from adding (h₂) alone. Please explain the differences between the main and controlled settings.

A1: Thank you for pointing this out. We apologize for not updating the experimental results in Table 3. We have updated the table and its discussion with the final results. All rows follow the inference procedure in Section 4.2. Complete BCR achieves 49.29% Rank-1 and 47.70% mAP, consistent with Table 1.

Adding h₂ raises Rank-1 from 49.21% to 49.29% and mAP from 47.67% to 47.70%, with the encoder, BFC, gallery graph, input scores and interval ordering held fixed. Complete RCR improves Rank-1 and mAP by +0.62% and +1.60%, respectively, while preserving Rank-5 and Rank-10.

Table A: Updated RCR component results on AERI-PEDES (%).

h₁h₂Interval orderingRank-1Rank-5Rank-10mAP
×××48.6766.1174.1746.10
✓✓×49.2966.1174.1746.18
××✓48.6766.1174.1747.61
✓×✓49.2166.1174.1747.67
✓✓✓49.2966.1174.1747.70

The last two rows isolate h₂: both retain h₁ and interval ordering, so only the second correction changes. The selection-only and ordering-only rows separate the two roles of RCR. Selection raises Rank-1, while interval ordering improves the positions of relevant images beyond the first result. Rank-5 and Rank-10 remain 66.11% and 74.17% in every row because the candidate sets are preserved.

The revised Table 3 and Section 4.4 now reflect these updated results.

W2: The benefit of reciprocal neighbors needs a direct comparison. RCR only keeps neighbors that select each other. The main paper does not compare this choice with ordinary nearest neighbors. A comparison using the same image features would help show whether reciprocal neighbors provide better support for ranking.

A2: Following your suggestion, we compared the four neighborhood rules in Table B on identical features, with the selection depth, correction bounds and intervals fixed. Each rule includes the candidate itself; the neighbor-based rules start from the same four nearest neighbors. Weighted 4NN uses a softmax of cosine similarities with temperature 0.1. We evaluated both fixed scorers and scorers refitted for each rule with four training seeds, keeping the encoder, BFC and training budget fixed.

The fixed-scorer comparison measures the effect of changing neighborhood support on the same trained representation and scoring functions. Refitting both scorers for each rule then tests the rule with coefficients learned for that neighborhood. These refits use training-order seeds 1, 1901, 1902 and 1903. The two comparisons therefore examine both the direct effect of the support rule and its behavior after matched scorer training.

Table B: Neighborhood comparison (%). Refitted results are mean ± sample standard deviation over four seeds. Link purity is the same-identity fraction of non-self links; labels are used only for this analysis.

DatasetNeighborhood ruleLink purityRank-1 (fixed scorers)mAP (fixed scorers)Rank-1 (refitted scorers)mAP (refitted scorers)
AERI-PEDESSelf only—48.6946.09——
AERI-PEDESOrdinary 4NN85.5649.1147.6049.646 ± 0.03647.682 ± 0.004
AERI-PEDESWeighted 4NN85.5649.7247.2249.687 ± 0.02047.212 ± 0.004
AERI-PEDESReciprocal 4NN (ours)95.6749.2947.7049.279 ± 0.01647.696 ± 0.002
TBAPRSelf only—49.1943.77——
TBAPROrdinary 4NN81.9149.8445.1950.094 ± 0.00045.182 ± 0.000
TBAPRWeighted 4NN81.9149.9144.5150.250 ± 0.00044.483 ± 0.001
TBAPRReciprocal 4NN (ours)88.8549.7244.8749.750 ± 0.03644.874 ± 0.003

With fixed scorers, all three neighbor-based rules improve Rank-1 and mAP over self-only support on both datasets. Reciprocity raises link purity from 85.56% to 95.67% on AERI-PEDES and from 81.91% to 88.85% on TBAPR. It gives the highest AERI-PEDES mAP with both fixed and refitted scorers; ordinary 4NN gives higher TBAPR mAP. These comparisons identify the contribution of gallery context and the role of reciprocity in selecting more identity-consistent neighbors. We added them to Appendix G.3.

AuthorsOpenReview ↗

Response to Reviewer TTQW

W3: The parameter choices need more support. The main paper does not clearly explain how the neighborhood size and the weight of neighborhood support are chosen. These choices may affect the final ranking. A small test with a few nearby values would help show whether the improvements hold across different settings.

A3: Thank you for this suggestion. Both datasets use k = 4 and α = 0.25. The small neighborhood keeps support local. In interval ordering, α gives neighbor support a weight of 25% and the image's own score a weight of 75%. Since this ordering follows first-result selection, α does not affect Rank-1.

Every tested k improves mAP on both datasets. Rank-1 also improves for all tested k on TBAPR and for k ≤ 6 on AERI-PEDES. Increasing α from 0 to 0.25 gives mAP gains of +1.51% and +1.04%, respectively. Nearby settings retain the gains, supporting the shared defaults.

Table C: Sensitivity to k and α (%), varying one parameter at a time with model weights fixed.

SettingAERI-PEDES Rank-1AERI-PEDES mAPTBAPR Rank-1TBAPR mAP
Direct ranking (no RCR)48.6746.1049.5643.79
k = 248.8747.0249.6644.37
k = 349.1347.4349.7244.64
k = 4, α = 0.25 (default)49.2947.7049.7244.87
k = 549.1347.7449.7245.10
k = 648.8447.7249.7545.15
k = 848.5147.4849.7545.26
α = 049.2946.1849.7243.83
α = 0.1049.2946.9249.7244.29
α = 0.2049.2947.4649.7244.70
α = 0.3049.2947.8949.7245.00
α = 0.4049.2948.2349.7245.26
α = 0.5049.2948.5049.7245.47

At α = 0, the trained first-result selection is still active; only neighborhood support in interval ordering is removed. This explains why that row differs from direct ranking and why Rank-1 stays fixed throughout the α sweep. Table C reports all tested values, with one parameter changed at a time and no retraining. Gains use unrounded metrics, while table entries are rounded for display. Appendix H includes the full sensitivity analysis.

W4: The effect of extra training is not clear. BCR starts from a CFAN checkpoint and trains the encoder for five more epochs. Table 2 does not report CFAN with the same extra training but without ADS, BFC, or RCR. Please add this baseline using the original CFAN training loss. This would help show how much of the improvement comes from extra training and how much comes from the proposed components.

A4: Thank you for suggesting this control. We continued each CFAN checkpoint for five epochs using its original objective (CDA + 0.5 FTA), matching the primary data order and optimization settings of encoder adaptation. The control uses no ADS, BFC or RCR and is evaluated at the fifth epoch. Table D compares it with the initial CFAN, the ADS-adapted encoder and complete BCR.

Table D: Matched five-epoch controls (Rank-1 / mAP, %). All entries are re-evaluated with the same inference procedure; Table 1 retains published baseline results. The Δ columns subtract initialization from control and control from BCR, using unrounded metrics.

DatasetCFAN initializationCFAN + 5 epochs (control)ADS encoder onlyComplete BCRΔ continued trainingΔ BCR vs. control
AERI-PEDES47.45 / 45.4448.05 / 45.8847.99 / 46.0849.29 / 47.70+0.60 / +0.45+1.24 / +1.81
TBAPR49.38 / 44.0249.63 / 44.1849.22 / 44.1149.72 / 44.87+0.25 / +0.16+0.09 / +0.69

The initialization-to-control comparison measures the effect of five additional epochs of the original CFAN objective. The control-to-BCR comparison then measures the gain beyond that continuation. The ADS-only column shows the adapted encoder before BFC and RCR are applied; the complete BCR column evaluates the full method. This separates additional primary training, auxiliary encoder supervision and the subsequent refinement stages within the same evaluation procedure.

Complete BCR improves both metrics over the control on both datasets. Rank-1/mAP gains are +1.24%/+1.81% on AERI-PEDES and +0.09%/+0.69% on TBAPR, showing improvement beyond additional primary training. The control matches primary updates; ADS adds an auxiliary batch every eight updates within the same optimizer step.

Across three paired training repeats, the AERI-PEDES mAP change from ADS is +0.173 ± 0.048% (mean ± sample standard deviation). These repeats complement the single-run controls in Table D. We added these controls to Section 4.6 and Appendix G.1, with the paired ADS comparison in Appendix C.1.

AuthorsOpenReview ↗

Response to Reviewer TTQW

Questions For Authors

Please address the weaknesses listed above. I believe these issues can all be resolved through further explanation and additional experiments. I am generally in favor of accepting the paper and hope the authors can address these concerns in their response.

Response: Thank you for your positive assessment. We have addressed the four concerns above through updated results, neighborhood comparisons, parameter sensitivity analyses and the matched extra-training control.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

We sincerely thank the Area Chair and all reviewers for their time, careful assessment, and constructive feedback. We have carefully considered each comment and provided detailed, point-by-point responses, supported by additional experiments and analyses.

We have revised the manuscript to present our method and findings more clearly. The additional evidence further supports the effectiveness of BCR and clarifies its contributions. We appreciate the opportunity to address the reviewers’ concerns and strengthen our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.