acceptodds
Under review as a conference paper at ICLR 2027

GPS: Geometry-Guided Patch Synthesis for Aerial-Ground Person Re-Identification

Abstract

Aerial-ground person re-identification must match identities despite substantial changes in local appearance and visibility across viewpoints. Existing methods primarily modify or align observed images, but do not explicitly construct identity-preserving cross-view variations with controllable matching difficulty. In this paper, we propose Geometry-Guided Patch Synthesis (GPS), a feature-level hard-positive learning framework that synthesizes structured changes in local identity evidence during training. Specifically, a Geometry-Constrained Patch Synthesizer (GCPS) module uses a fixed spatial coordinate prior to reorganize patch tokens through three complementary operations: compression, merging, and vanishing, thereby producing candidates with controlled local variations. Then, a Hard-Positive Patch Synthesis (HPPS) module is proposed to select the hardest identity-consistent candidate within a prescribed difficulty range relative to a real same-identity cross-view observation. Finally, a Dual-Path Evidence Fusion (DPEF) module transfers locally supported correspondence between the original and selected synthetic features into identity learning while preserving unsupported original evidence. Extensive experiments on four aerial-ground benchmarks demonstrate the effectiveness of GPS, with consistent improvements over Plain CLIP across all evaluated protocols. Component ablations further demonstrate the complementary contributions of structured patch synthesis, difficulty-controlled selection, and evidence fusion. The code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aerial-Ground Person Re-Identification, Geometry-Guided Patch Synthesis, Hard-Positive Learning, Cross-View Representation Learning, Evidence Fusion

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 4 reviews, above 85% of the others.
Reviewer_PWCyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GPS, a hard-positive learning framework that addresses the insufficient modeling of identity-preserving cross-view variations in aerial-ground person re-identification. GPS first employs GCPS to generate cross-view candidates with varying strengths, then uses HPPS to select the hardest candidate within a prescribed difficulty range while satisfying identity constraints. Finally, DPEF fuses supported evidence from the original and selected candidate features to improve identity representation learning.

Critical strengths
  1. The paper models aerial-ground viewpoint differences as changes in the spatial distribution, relative contribution, and visibility of local identity evidence, which is intuitive and well aligned with practical AG-ReID scenarios.
  2. Using real same-identity cross-view pairs as references, the method generates identity-consistent candidates with controllable additional difficulty, making it more targeted than unconstrained data augmentation.
  3. The appendix provides detailed method parameters, training configurations, and evaluation protocols, offering substantial information for reproducibility.
  4. The synthesis, selection, and fusion modules are used only during training. Inference still requires a single image encoding followed by cosine-distance retrieval, introducing no additional inference overhead.
Critical weaknesses
  1. About GCPS

Appendix B.1 explicitly defines the coordinate prior zi, but this prior is essentially a fixed grid function and does not incorporate actual camera angles or spatial coordinates. It implicitly assumes that the person is roughly centered and that the body axis is vertical; tilted poses or detection-box offsets may violate these assumptions.

Merging and vanishing treat the min–max normalized token norm as a local response and use it to compute support. Although intuitively reasonable, this assumption is not directly validated. Moreover, the vanishing risk function tends to select tokens with lower positional support and weaker responses. This creates some tension with its motivation of modeling cases where “distinctive details become invisible” (under this proxy, salient cues with stronger responses are instead more likely to be retained).

  1. About HPPS

HPPS computes Δk using a global representation h that combines the class token and pooled patch information. Local edits affecting only a small region may be diluted during feature aggregation, and the paper does not verify whether this global difficulty measure accurately reflects the strength of local edits.

The design of HPPS is intuitively reasonable, but the paper provides no statistics or analysis of its actual selection behavior, such as the candidate acceptance rate or the proportion of samples with no accepted candidate. It is therefore difficult to determine whether the module indeed selects candidates that are moderately difficult and identity-consistent.

  1. About experiments

All comprehensive ablations are conducted on the synthetic CARGO dataset, which is insufficient to establish the independent contribution and stability of each component under real-world capture conditions.

The paper lacks a generic augmentation baseline under matched conditions. For example, if random token perturbations combined with HPPS and DPEF yield comparable improvements, the main gains may come from appropriate hard samples and additional supervision rather than the geometry-guided design.

Questions for authors
  1. How robust is the fixed coordinate prior in GCPS to tilted poses and off-center detections? Could the authors provide analyses under these conditions?

  2. Is there empirical evidence that the min–max normalized token norm reliably reflects local identity evidence or support? In particular, since the vanishing risk function favors tokens with lower positional support and weaker responses, how does it model the disappearance of distinctive, highly salient identity cues?

  3. How well does the global difficulty measure Δk, computed from the class token and pooled patch information, capture changes in matching difficulty caused by edits affecting only a small number of tokens? Could the authors validate its sensitivity to localized changes?

  4. Could the authors report statistics on the actual behavior of HPPS, such as the candidate acceptance rate and the proportion of samples with no accepted candidate, as well as the distributions of selected candidate strengths and Δk?

  5. Could the authors provide component ablations on at least one dataset with real captured images to verify that the individual contributions and stability observed on synthetic CARGO data generalize to real-world capture conditions?

  6. How does GPS compare with generic feature-space augmentation under matched conditions? For example, could the authors evaluate random token perturbation, masking, or mixing using the same number of candidates, HPPS, DPEF, and training objectives, to isolate the contribution of the geometry-guided design?

8 comments

We sincerely thank you for your thorough review and suggestions.

Q1: (a) Appendix B.1 explicitly defines the coordinate prior ziz_i, but this prior is essentially a fixed grid function and does not incorporate actual camera angles or spatial coordinates. It implicitly assumes that the person is roughly centered and that the body axis is vertical; tilted poses or detection-box offsets may violate these assumptions.

(b) How robust is the fixed coordinate prior in GCPS to tilted poses and off-center detections? Could the authors provide analyses under these conditions?

A1:

(a)

First, person ReID typically follows pedestrian detection. The detector provides a bounding box, and the ReID model receives a crop around the person. This makes the person's location and body layout relatively consistent in the model input. We also observed this pattern across several real-capture benchmarks. We manually reviewed a stratified random sample of 360 training crops. A crop was compatible if the principal body axis was within 30° of vertical and the horizontal and vertical offsets of the center of the body's extent from the crop center were at most 20% of the crop width and height, respectively. Most training crops were compatible with the prior, with particularly high compatibility rates on real-capture benchmarks.

Table K: Manual audit of crop layouts.

DatasetCompatible crops
AG-ReID v189.6%
AG-ReID v287.4%
LAGPeR95.8%
CARGO66.9%

Second, explicit viewpoint projection and scale modeling require richer per-observation information, such as camera intrinsics, viewing angle, and camera pose or height. GPS is designed to use the person crops and aerial–ground view labels already provided by standard AG-ReID benchmarks without requiring additional annotations. In future work, we will further explore using actual camera geometry for finer geometric modeling.

Finally, the fixed prior provides a shared coordinate reference and serves only as soft supervision. The locations of actual edits depend on the input image's patch response strengths and vary across samples. The crop center serves as a positional reference, and patch features can match across grid locations. This lets cues from tilted or off-center pedestrians participate in matching where they appear. For example, if a shifted detection box places a shoulder cue to one side of the crop, its token response features can still establish a correspondence with a donor patch.

(b)

We applied random rotation, horizontal translation, and vertical translation separately, then retrained and evaluated GPS, ViSA, and GSAlign under matched conditions. Relative to the original results, GPS scores dropped by about 1–2 percentage points, compared with about 3–5 points for ViSA and GSAlign. GPS still achieved the highest mAP and Rank-1 on both AG-ReID v2 protocols under each matched perturbation. The smaller drops indicate greater robustness to tilted poses and off-center crops. We believe the geometric prior and GCPS's use of input token responses help the method adapt to these changes.

Table K1: Rotation and translation robustness on AG-ReID v2 (mAP / Rank-1, %).

SettingViSA A→CViSA C→AGSAlign A→CGSAlign C→AGPS A→CGPS C→A
Original (Table 1(b))83.61 / 89.4382.32 / 88.5781.38 / 87.8681.05 / 88.0284.75 / 89.9083.58 / 89.01
Random rotation80.14 / 85.3178.06 / 84.9977.20 / 84.3077.66 / 83.4083.31 / 88.4282.04 / 87.81
Horizontal translation78.97 / 86.0778.81 / 83.8477.86 / 83.1576.48 / 84.5483.57 / 88.1181.83 / 87.62
Vertical translation79.78 / 84.6277.60 / 85.3076.69 / 84.6277.19 / 83.8383.06 / 88.6582.37 / 87.18

Q2: (a) Merging and vanishing treat the min–max normalized token norm as a local response and use it to compute support. Although intuitively reasonable, this assumption is not directly validated. Moreover, the vanishing risk function tends to select tokens with lower positional support and weaker responses. This creates some tension with its motivation of modeling cases where “distinctive details become invisible” (under this proxy, salient cues with stronger responses are instead more likely to be retained).

(b) Is there empirical evidence that the min–max normalized token norm reliably reflects local identity evidence or support? In particular, since the vanishing risk function favors tokens with lower positional support and weaker responses, how does it model the disappearance of distinctive, highly salient identity cues?

A2: To illustrate how local responses inform support, we visualized patch-token responses. Please refer to Figure R1 in §4.9 of the latest version of the paper. The top row shows the normalized token-response heatmaps, their overlays, and the 26 highest-response tokens (Top26). High responses are concentrated mainly on the pedestrian's body.

Some low-response patches on the body still contain subtle appearance details. In the figure, some low-response patches appear near the hands, waist, or feet or in the background.

In addition, we compared three response definitions on AG-ReID v2.

  • Position prior only. Use positional support ziz_i alone to guide the edits.
  • Quantile-clipped normalization. Clip the tails of the within-image token-norm distribution at quantile bounds, then rescale the clipped norms.
  • Min–max token norm (GPS). Apply min–max normalization to the raw L2L_2 norms of the current patch sequence.

The table reports mAP / Rank-1 (%) on the four AG-ReID v2 protocols.

Table L: Comparison of token-response definitions on AG-ReID v2.

Response definitionA→CC→AA→WW→A
Position prior only81.72 / 87.6980.68 / 87.0885.55 / 90.7681.54 / 87.61
Quantile-clipped normalization82.35 / 88.0381.22 / 87.6386.18 / 91.4982.37 / 88.59
Min–max token norm (GPS)84.75 / 89.9083.58 / 89.0188.29 / 93.3585.11 / 90.98

The min–max token norm improves both mAP and Rank-1 over both alternatives on all four protocols, with mAP gains of 2.11 to 2.74 percentage points over quantile-clipped normalization. These results support using the within-image min–max norm as the local response for patch editing.

Vanishing conservatively attenuates local details after compression and merging. It selects a small set of regions with weaker token responses to simulate the partial loss of local evidence. Suppressing the most discriminative regions can make a hard positive unreliable as a same-identity sample. We will revise the phrase “distinctive details” in the paper to avoid implying that vanishing suppresses salient cues. We also tried prioritizing high-response tokens for vanishing, which made candidates too difficult and led to the loss of salient identity cues. On all four AG-ReID v2 protocols, prioritizing high-response tokens gives lower mAP and Rank-1 than the original GPS. The table reports mAP / Rank-1 (%).

Table M: Comparison of vanishing high- and low-response tokens on AG-ReID v2.

Vanishing settingA→CC→AA→WW→A
Prioritize high-response tokens81.16 / 87.2779.84 / 86.8684.24 / 90.4081.77 / 88.55
Original GPS (prioritize low-response tokens)84.75 / 89.9083.58 / 89.0188.29 / 93.3585.11 / 90.98

In addition, with compression and merging fixed, adding vanishing on the real-capture LAGPeR G→AG protocol raises mAP / Rank-1 from 25.91 / 33.55 to 27.85 / 36.18 (A5 below). The ablation gain also supports preferentially weakening a small set of low-response tokens after compression and merging, with positional support guiding the selection.


Q3: (a) HPPS computes Δk using a global representation h that combines the class token and pooled patch information. Local edits affecting only a small region may be diluted during feature aggregation, and the paper does not verify whether this global difficulty measure accurately reflects the strength of local edits.

(b) How well does the global difficulty measure Δk, computed from the class token and pooled patch information, capture changes in matching difficulty caused by edits affecting only a small number of tokens? Could the authors validate its sensitivity to localized changes?

A3: Retrieval uses global descriptors, so local edits aim to make same-identity matching harder at the level of the global representation. Changes in global similarity show whether the edits affect retrieval. We used two CARGO diagnostics to examine edit strength and extent.

First, we compared unedited features with candidates generated at five strengths for the same 960 CARGO sources. We scored the candidates generated by full GPS. The table reports mean Δk\Delta_k in units of 10−310^{-3}.

Table N: Mean added difficulty by synthesis strength on CARGO.

Generation level ℓMean Δk\Delta_k
No edit≈0
0.35003.096539
0.51253.862922
0.67504.645124
0.83755.443020
1.00006.256495

The unedited control has a mean Δk\Delta_k of approximately zero. With editing, the mean is positive at all five strengths and increases as the edits get stronger. This positive relationship indicates that stronger patch edits increase global matching difficulty, yielding harder same-identity positives.

Second, we held the CLS token and donor fixed and kept the actual GCPS edit residuals at a specified number of source patch positions. The multiplier α\alpha controls the magnitude of the retained residuals. The table reports mean Δk\Delta_k in units of 10−510^{-5}.

Table O: Mean added difficulty by edited-token count and residual scale on CARGO.

Edited tokens out of 128α = 0.25α = 0.50α = 1.00
10.1800.1890.218
40.4910.7820.916
162.4163.1516.732
326.31610.06721.663
6416.72434.345101.650
12871.691260.5711105.382

Within each column, at fixed α\alpha, the difficulty increment increases as more tokens are edited. Within each row, at a fixed number of edited tokens, stronger edits also increase the difficulty increment. The global difficulty measure Δk\Delta_k thus captures the extent and strength of local edits.


Q4: (a) The design of HPPS is intuitively reasonable, but the paper provides no statistics or analysis of its actual selection behavior, such as the candidate acceptance rate or the proportion of samples with no accepted candidate. It is therefore difficult to determine whether the module indeed selects candidates that are moderately difficult and identity-consistent.

(b) Could the authors report statistics on the actual behavior of HPPS, such as the candidate acceptance rate and the proportion of samples with no accepted candidate, as well as the distributions of selected candidate strengths and Δk?

A4: We analyzed HPPS selection behavior on CARGO. In the training logs, every batch contains 64 original benchmark images, and 100% of sources have an observed same-identity cross-view reference. During the active synthesis stage, 75% of eligible sources enter GCPS. GCPS generates five candidates for each selected source, and HPPS accepts at most one. The recorded mean source-level acceptance is 79.88%, corresponding to estimated coverage of 75% × 79.88% = 59.91%, or about 38 of 64 original inputs receiving an accepted synthetic-feature update. These updates add cross-view evidence beyond the original observations used for training.

In the candidate-level analysis, approximately 25.5501% of candidates passed all gates, 77.3456% of participating sources had at least one accepted candidate, and 22.6544% fell back to their original features. Among the selected candidates, Δₖ had a mean of 0.0050357, a median of 0.0048971, and a 5th–95th percentile interval of [0.0040995, 0.0064198] within the default [0.004, 0.007] band.

We also tallied the shares of accepted synthetic samples at each nominal strength, with 953 recorded selections in total.

Table P: Distribution of HPPS selections by nominal synthesis strength on CARGO.

Nominal strength ℓSelectionsShare of selections
0.350010.1049%
0.5125313.2529%
0.675018118.9927%
0.837537839.6642%
1.000036237.9853%

The two highest nominal strengths account for 77.65% of the selections. This pattern is consistent with the Q3 results, where the two lowest strengths produce mean Δₖ values below the 0.004 lower bound, and with HPPS choosing the hardest candidate that satisfies the identity and realized difficulty checks.


Q5: (a) All comprehensive ablations are conducted on the synthetic CARGO dataset, which is insufficient to establish the independent contribution and stability of each component under real-world capture conditions.

(b) Could the authors provide component ablations on at least one dataset with real captured images to verify that the individual contributions and stability observed on synthetic CARGO data generalize to real-world capture conditions?

A5: We ran the operator, module, and training objective ablations on all four datasets while preparing the paper. Because of space constraints, Table 2 in the main text presents only the CARGO ablations. We will add the complete ablation results for all four datasets to the appendix. The tables below report all LAGPeR and AG-ReID v1 protocols and two AG-ReID v2 protocols. The controls follow the settings in Table 2 and Appendix C.3. Entries are mAP / Rank-1 (%).

LAGPeR

Table Q: Ablations on LAGPeR.

Ablation settingA→GG→AG→AG
Operators
No C, M, or V29.86 / 42.0933.06 / 34.3721.45 / 27.91
C32.61 / 44.8535.90 / 37.9824.56 / 31.94
M32.37 / 45.0435.50 / 37.4623.87 / 31.55
V31.44 / 43.6334.69 / 36.5723.41 / 30.60
C + M33.94 / 46.8537.04 / 39.3025.91 / 33.55
C + V34.14 / 46.8837.39 / 39.7226.20 / 33.75
M + V33.60 / 46.2937.06 / 39.3025.51 / 33.22
C + M + V35.67 / 48.5639.02 / 41.6327.85 / 36.18
Modules
Clean control29.20 / 40.7732.39 / 34.2421.13 / 27.58
GCPS34.53 / 46.2937.71 / 40.1526.35 / 34.41
GCPS + HPPS35.37 / 48.1638.71 / 41.3327.52 / 35.72
GCPS + DPEF35.17 / 48.1338.35 / 41.0027.39 / 35.65
GCPS + HPPS + DPEF35.67 / 48.5639.02 / 41.6327.85 / 36.18
Objectives
No GVPS or OSIC34.17 / 46.9537.41 / 39.8226.22 / 33.85
GVPS34.75 / 47.1438.18 / 40.6427.02 / 35.26
OSIC34.97 / 47.7038.33 / 40.6827.24 / 35.33
GVPS + OSIC35.67 / 48.5639.02 / 41.6327.85 / 36.18
Reviewer_7HHeOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Geometry-Guided Patch Synthesis (GPS), a feature-level hard-positive learning framework for aerial-ground person re-identification (AG-ReID). The method aims to address substantial viewpoint-induced appearance variations by constructing identity-preserving synthetic features with controlled matching difficulty. Specifically, the Geometry-Constrained Patch Synthesizer (GCPS) models local appearance variations through compression, merging, and vanishing operations. Hard-Positive Patch Synthesis (HPPS) selects challenging yet identity-consistent candidates based on their relative difficulty compared with real cross-view pairs. Dual-Path Evidence Fusion (DPEF) further integrates the complementary information from original and synthesized features to improve identity representation learning. The authors conduct extensive experiments on four benchmarks, including AG-ReID v1, AG-ReID v2, CARGO, and LAGPeR. The results demonstrate consistent improvements over the Plain CLIP baseline across the evaluated protocols. Additional ablation studies, parameter analyses, and experiments with different backbones and ReID frameworks further support the effectiveness and applicability of the proposed method.

Critical strengths

1.The paper provides an interesting perspective on aerial-ground person re-identification by treating viewpoint-induced appearance changes as variations in the spatial organization, contribution, and visibility of local identity evidence. 2.The proposed framework is technically reasonable, and its main components are supported by mathematical formulations and experimental evaluations. The ablation studies provide evidence for the complementary contributions of the proposed modules.

Critical weaknesses

Although the three patch operations are motivated by aerial-ground viewpoint variations, the connection between the fixed spatial prior and the actual geometric changes could be discussed in greater detail. A brief explanation of why the selected geometric formulation is appropriate for different viewpoints would make the motivation more convincing.

Questions for authors

Q1. The proposed GCPS employs a fixed spatial coordinate prior to guide patch compression, merging, and vanishing. Could the authors provide some additional intuition on how this prior accommodates different aerial-ground viewpoints and variations in pedestrian appearance? Q2. HPPS selects hard-positive candidates according to their difficulty relative to real cross-view pairs. Could the authors briefly elaborate on the advantages of this pair-relative difficulty measure compared with directly selecting candidates based on their absolute feature distances? Q3. DPEF distinguishes shared evidence, original-only evidence, and synthetic-only evidence during feature fusion. Could the authors provide further intuition on how this distinction contributes to preserving discriminative identity information, particularly when local appearance changes are substantial? Q4. Table 3 demonstrates consistent improvements across several backbone architectures and ReID frameworks. Could the authors briefly discuss why the performance gains vary across these models and whether the observed differences are related to their underlying feature representations?

5 comments

We sincerely thank you for your thorough review and suggestions.

Q1: (a) Although the three patch operations are motivated by aerial-ground viewpoint variations, the connection between the fixed spatial prior and the actual geometric changes could be discussed in greater detail. A brief explanation of why the selected geometric formulation is appropriate for different viewpoints would make the motivation more convincing.

(b) The proposed GCPS employs a fixed spatial coordinate prior to guide patch compression, merging, and vanishing. Could the authors provide some additional intuition on how this prior accommodates different aerial-ground viewpoints and variations in pedestrian appearance?

A1: The fixed grid provides a soft spatial reference. Patches can still match across grid locations when the detection box shifts, and merging weights and vanishing locations depend on the input token responses. Consider a pedestrian wearing a jacket with a chest logo and a black shoulder strap across the body.

  • Compression models changes in the spacing between cues. Viewed horizontally from the ground, the chest logo and jacket hem may be several patch rows apart. Viewed from above, their vertical separation may shrink. Compression resamples along the same column so the cues occupy a shorter vertical span while retaining their top-to-bottom order.
  • Merging models changes in the relative prominence of nearby cues. From another angle, the shoulder strap and fabric texture may overlap within a small region. The strap's outline can remain clear while the fabric texture fades. Merging lets tokens with stronger local support retain more of their own features, while those with weaker support incorporate more neighborhood information.
  • Vanishing models the loss of local details. From yet another angle, only part of a small sleeve badge near the edge of the body may remain visible, making it hard to recognize. Vanishing prioritizes tokens with lower positional support and weaker local responses for context replacement, simulating details that become unavailable.

Our supplementary manual audit found that most crops were compatible with the prior, particularly on real-capture benchmarks. We manually reviewed 360 stratified random training crops. The principal body axis had to be within 30° of vertical and the center of the body's extent within 20% of the crop width and height from the crop center along the respective axes.

Table I: Manual audit of crop-layout compatibility across benchmarks.

DatasetCompatible crops
AG-ReID v189.6%
AG-ReID v287.4%
LAGPeR95.8%
CARGO66.9%

The fixed geometric prior accommodates changes in aerial–ground viewpoint and pedestrian appearance by providing a common coordinate reference. It serves only as soft supervision. The locations selected for editing depend on patch-response strength and vary across inputs.


Q2: HPPS selects hard-positive candidates according to their difficulty relative to real cross-view pairs. Could the authors briefly elaborate on the advantages of this pair-relative difficulty measure compared with directly selecting candidates based on their absolute feature distances?

A2: HPPS controls how much matching difficulty a synthetic candidate adds relative to its current observed cross-view positive pair. Equation 5 defines this increment as Δ_k = d(h_k, h_d) − d(h_o, h_d), with d(a, b) = 1 − cos(a, b). The first term is the distance from the synthetic candidate to the observed donor. The second is the distance within the observed cross-view positive pair.

Here h_o and h_d are the observed source and donor representations of the same identity across views, and h_k is the synthetic candidate representation. Viewpoint, occlusion, and image clarity can make observed pairs differ substantially in their starting difficulty.

Observed pair distanceCandidate-to-donor distanceAdded difficulty Δ_k
0.1000.1060.006
0.3000.3060.006

The absolute distances differ substantially. Both edits add 0.006. A shared interval can therefore require each candidate to be moderately harder than its own observed pair. This lets a candidate from an already difficult pair qualify when its added difficulty is within the band.

  • Calibration to the current representation. The observed distance evolves during training. Subtracting the current source–donor distance reduces the effect of that shift on acceptance.
  • Control over the added difficulty. The lower bound requires a meaningful increase in difficulty, and the upper bound caps that increase. A separate gate checks identity consistency (Appendix B.2).

For a given source and donor, HPPS subtracts the same observed distance from all candidates, so the ranking is unchanged. The benefit is the calibration of the acceptance interval, which makes the added difficulty comparable across observed pairs.

Q3: DPEF distinguishes shared evidence, original-only evidence, and synthetic-only evidence during feature fusion. Could the authors provide further intuition on how this distinction contributes to preserving discriminative identity information, particularly when local appearance changes are substantial?

A3: When local appearance changes substantially, a synthetic observation can still preserve the person's identity while correspondence reliability varies by position. DPEF distinguishes three evidence states at each position. Consider a jacket with a black shoulder strap and chest logo in the original image. After synthesis, the strap shifts slightly, the logo becomes blurred, and a conspicuous bright stripe appears nearby.

  • The shoulder strap is shared evidence. A reliable correspondence remains even after the strap shifts. DPEF uses the aligned synthetic feature for a weighted update, helping the model recognize the same cue after the shift.
  • The chest logo is original-only evidence. Reliable correspondence is difficult to find after synthesis, while the logo in the original image remains clear and helps distinguish people wearing similar jackets. DPEF preserves this cue in the original path with a small adjustment.
  • The bright stripe is synthetic-only evidence. The stripe is salient in the synthetic features and has weak correspondence support in the original image. DPEF excludes it from the local residual added to the original features.

Larger local appearance changes make it more likely that some cues can still be matched while others remain clear only in the original image. Distinguishing these three evidence states lets the model learn from the shoulder strap's shift, preserve discriminative details such as the chest logo, and limit the influence of newly synthesized textures on local fusion.

The three states let DPEF learn correspondences that survive appearance changes and preserve discriminative details that remain clear in the original observation (Equations 6–7 and Appendix B.3). With GCPS and HPPS fixed, adding DPEF improves CARGO A→G mAP from 72.35% to 73.35% and Rank-1 from 73.40% to 74.47% (Table 2(b)). These gains come from evidence fusion within the same synthesis and selection pipeline.


Q4: Table 3 demonstrates consistent improvements across several backbone architectures and ReID frameworks. Could the authors briefly discuss why the performance gains vary across these models and whether the observed differences are related to their underlying feature representations?

A4: The gains in Table 3 and the supplementary integrations can be understood by considering how much local variation GPS adds where the architecture has limited coverage and how directly the architecture can use it for retrieval.

GSAlign uses learnable thin-plate spline transformations and channel-wise weighting of class prototypes to align observed cues. GPS complements this by exposing the model to same-identity examples in which local cues contract, mix, or become weaker. For example, if a cuff mark becomes faint in another view, the model can learn to match the identity using the remaining cues. This complementarity is consistent with GSAlign showing the largest gains, +10.87 percentage points in mAP and +11.54 in Rank-1.

VDT and SeCap share a VDT encoder design. VDT mainly retrieves with its global meta token. SeCap also uses separately supervised local descriptors in its retrieval representation, giving patch-level training signals a more direct route to the final ranking. Their mAP gains are +4.04 and +8.15 points, respectively.

ViSA and HiHR already model some of the local variation that GPS targets. ViSA uses expert-generated queries and local graph aggregation to select cues available under different views. HiHR fuses features from multiple layers and uses coarse-to-fine supervision to capture shared cross-view cues and view-specific details. GPS adds supervision that partly overlaps with these mechanisms, consistent with the more moderate mAP gains of +3.44 points for ViSA and +3.59 for HiHR.

The gains can be especially pronounced at the top of the ranking. ViSA gains 6.41 points in Rank-1. DINOv3, whose backbone provides dense local features, gains 8.24 points in Rank-1 and 3.82 in mAP. Table 3 reports improvements in both metrics for all six paired configurations, including +1.66 mAP and +2.24 Rank-1 points for TransReID.

Table J: Paired GPS integration results for five dedicated ReID frameworks on CARGO ALL.

FrameworkBase mAP / Rank-1 (%)With GPS mAP / Rank-1 (%)Gain (percentage points)
VDT (Table 3)55.20 / 64.1059.24 / 66.71+4.04 / +2.61
SeCap (Table 3)60.19 / 68.5968.34 / 72.12+8.15 / +3.53
ViSA65.46 / 70.5168.90 / 76.92+3.44 / +6.41
HiHR67.52 / 74.3671.11 / 76.28+3.59 / +1.92
GSAlign57.95 / 65.0668.82 / 76.60+10.87 / +11.54

Dear Reviewer,

Thank you for taking the time to review our paper and for your constructive feedback. We believe our responses address the concerns you raised.

With fewer than two days remaining in the discussion period, we wanted to check whether you have any further questions or concerns. If there is anything you would like us to clarify, please feel free to let us know. We would be happy to respond promptly before the discussion closes.

Best regards, The Authors

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

This paper proposes Geometry-Guided Patch Synthesis (GPS) for aerial-ground person re-identification. The method constructs challenging positive examples in feature space to improve recognition under changes in local appearance and visibility. Its Geometry-Constrained Patch Synthesizer uses a fixed spatial prior to perform patch compression, merging, and vanishing, together with information from a same-identity cross-view donor. Hard-Positive Patch Synthesis selects candidates within a prescribed difficulty range while applying identity-consistency checks. Dual-Path Evidence Fusion then integrates supported correspondences between the original and synthetic features into training. The auxiliary synthesis and fusion components are removed at inference, which uses a single image encoder. Experiments on four benchmarks show consistent improvements over Plain CLIP. Additional evaluations examine component contributions, applicability across backbones, and cross-dataset generalization.

Critical strengths
  1. The motivation of this work is reasonable. The paper addresses changes in the spatial extent, relative contribution, and visibility of identity cues across aerial and ground views. Constructing difficult positive examples is a reasonable complement to aligning existing observations. Measuring additional difficulty relative to a real same-identity cross-view pair is also more informative than controlling augmentation strength alone.

  2. The performance of GPS is good. GPS consistently improves over Plain CLIP across the evaluated protocols and achieves competitive results against the listed methods. The improvements are substantial in several settings. For example, on LAGPeR under G→AG, Rank-1 increases from 26.30% to 36.18%. The evaluation covers both captured and synthetic imagery, rather than relying on a single dataset.

  3. The paper evaluates GPS with several backbones and ReID frameworks, including ViT, DINOv3, VDT, SeCap, and TransReID. Appendix A.2 states that each baseline/GPS pair shares the relevant training and evaluation settings. The cross-dataset evaluation from LAGPeR to AG-ReID v1/v2 further supports the transferability of the overall approach.

Critical weaknesses
  1. The presentation of the method could be clearer. The framework contains several modules and auxiliary steps, and some important details are deferred to Appendix B. Although Figure 2 provides an overview, a short algorithm showing candidate construction, selection, fusion, and supervision would make the training procedure easier to follow. The main text could also give slightly more space to explaining the roles of GVPS and OSIC.

  2. It is interesting to find that the compression can be used for both source platforms.. Either aerial or ground images can serve as the source, while compression is mainly motivated by local evidence becoming spatially concentrated under elevated viewpoints. The authors should clarify why the same compression operation is appropriate for both source platforms and whether its intended role differs between them.

  3. This is reasonable for constructing challenging positives in donor selection, but the authors should briefly discuss the trade-off between donor difficulty and reliability, particularly during early training epochs.

Questions for authors

Please see the weaknesses.

5 comments

We sincerely thank you for your thorough review and suggestions.

Q1: The presentation of the method could be clearer. The framework contains several modules and auxiliary steps, and some important details are deferred to Appendix B. Although Figure 2 provides an overview, a short algorithm showing candidate construction, selection, fusion, and supervision would make the training procedure easier to follow. The main text could also give slightly more space to explaining the roles of GVPS and OSIC.

A1: We will add a short algorithm to the main text to show how candidate construction, selection, fusion, and supervision fit together. We will also explain the respective roles of GVPS and OSIC there. Please refer to §3.4 and Algorithm 1 in the latest version of the paper.

Algorithm 1. GPS training step

Input   Batch B (images, IDs, views), model θ, progress p, queue Q
Output  Updated model θ

F ← Encode(B, θ)
A ← ∅
if SynthesisActive(p):
    D ← MatchDonors(F, B, Q)
    for (s, d) in TopGapPairs(D, 0.75):
        C ← [GCPS(s, d, p, ℓ) for ℓ in Levels(5)]
        C ← HPPS_HardenAndScore(C, s, d, p)
        c ← HPPS_Select(C)
        if c = ∅ and NoRawInBand(C) and HasOverHard(C):
            C_cal ← CalibrateAndScore(C, s, d)
            c ← HPPS_Select(C_cal)
        if c = ∅: continue
        r ← DPEFResidual(s, c)
        g ← BoundedUpdate(s, d, Pool(r))
        A ← A ∪ {(s, c, g)}

L ← L_base(B, F, A) + λ_G L_GVPS(A) + λ_O L_OSIC(A)
return OptimizerStep(θ, ∇_θ L)

MatchDonors chooses the least similar same-identity donor from the other platform, searching the current batch first and using the detached queue Q if no eligible donor is found. TopGapPairs keeps the 75% of eligible sources with the largest cosine gaps. GCPS uses compression, merging, vanishing, and bounded donor-conditioned edits to construct five candidates at the scheduled strengths.

HPPS_Select selects the candidate with the largest donor-relative Δₖ among those passing validity, identity-consistency, and difficulty-band checks. CalibrateAndScore calibrates an overly hard candidate toward the source features and recomputes its scores. A source with no accepted candidate keeps its clean representation. A collects the accepted sources, their selected candidates, and the global features updated with the pooled DPEF residual. L_base includes clean supervision and synthetic replay on accepted sources, which also receive GVPS and OSIC.

Consider a person wearing a striped jacket and carrying a shoulder bag. The ground crop is the source, and an aerial crop of the same person is the donor. In the aerial view, the jacket occupies a shorter vertical span. To model this change, GCPS compresses the stripe pattern into fewer token rows, adjusts the relative contributions of neighboring jacket and bag tokens, and attenuates a weak detail such as part of the bag strap. The aerial donor guides bounded changes to the source features.

Suppose the five candidates increase the cosine distance to the donor by 0.0020, 0.0045, 0.0060, 0.0075, and 0.0090 relative to the observed source–donor distance. Among candidates that pass the validity and identity checks, HPPS selects the 0.0060 candidate as the hardest one in the default band [0.004, 0.007]. DPEF can match a displaced stripe to a nearby original patch and use that supported correspondence in fusion. If the strap becomes faint in the selected candidate and its original patch retains the stronger response, DPEF preserves it as original-only evidence. The original crop retains its identity supervision. The selected candidate contributes replay supervision, with GVPS guiding the spatial coherence of the edits and OSIC aligning the original and synthetic identity representations.

The Geometry-Valid Patch Synthesis (GVPS) objective regularizes neighboring edits for spatial coherence and penalizes excessive feature shifts, low identity confidence, and weak correspondence support. The Original-Synthetic Identity Consistency (OSIC) objective aligns the clean and synthetic identity representations at the global level and at supported local positions. It also controls the difficulty of the final descriptor relative to the donor, so synthetic samples maintain identity consistency and continue to provide a useful hard-positive signal (Section 3.4 and Appendix B.4).

Q2: It is interesting to find that the compression can be used for both source platforms.. Either aerial or ground images can serve as the source, while compression is mainly motivated by local evidence becoming spatially concentrated under elevated viewpoints. The authors should clarify why the same compression operation is appropriate for both source platforms and whether its intended role differs between them.

A2: For ground source images, compression simulates the foreshortening induced by an elevated viewpoint, where vertically separated cues can become concentrated in a smaller area. For aerial source images, further concentrating local cues can produce harder positive samples that help the model learn to align aerial and ground features. We compared applying compression only to aerial sources, only to ground sources, and to both, as in GPS. The single-platform variants change only which sources receive compression. The two-platform setting gives the highest mAP and Rank-1 in all four protocols. This supports applying compression to both platforms and matches our intuition that compressing ground features and further compressing aerial features complement each other in cross-view training.

Table G: Effects of applying compression to aerial sources, ground sources, or both on CARGO and LAGPeR.

Compression applied toCARGO ALLCARGO A→GLAGPeR A→GLAGPeR G→A
Plain CLIP (Table 1)63.49 / 69.5563.63 / 65.9628.78 / 40.7431.91 / 33.13
Aerial sources only69.20 / 75.9670.77 / 72.3433.49 / 45.6037.13 / 39.72
Ground sources only70.24 / 76.2870.61 / 73.4034.31 / 47.2837.63 / 40.32
Aerial and ground sources (GPS)71.55 / 77.2473.35 / 74.4735.67 / 48.5639.02 / 41.63

Q3: This is reasonable for constructing challenging positives in donor selection, but the authors should briefly discuss the trade-off between donor difficulty and reliability, particularly during early training epochs.

A3: Early in training, the encoder is still learning to distinguish identities and has yet to learn stable feature representations, so donor rankings based on feature similarity can fluctuate. Enabling GPS at full strength at this stage does not produce useful hard positives. We first train the encoder on the original observations and enable GPS after it has learned an initial ReID representation. On CARGO, the GPS stage starts from a trained Plain CLIP ReID checkpoint. The default strength multiplier q(p) = clip(2p, 0, 1) rises from zero to one during the first half of the GPS stage (Appendix B.1).

Once synthesis is active, HPPS balances difficulty and reliability. A candidate that is too easy adds little training signal, and an overly difficult edit can weaken discriminative identity cues. The identity checks cover proximity to the source, correct identity prediction, and prediction confidence. The original representation is used when a source has no accepted candidate.

Our initial trials with GPS at full strength from the start gave lower retrieval scores, motivating the gradual schedule. The supplementary CARGO ALL comparison below reports mAP and Rank-1 (%). The ramp fraction τ denotes how much of the GPS stage is used to reach full strength. The default reaches full strength halfway through the stage.

Table H: Comparison of synthesis strength schedules on CARGO ALL.

Curriculumramp fraction τmAPRank-1
Fast ramp0.1063.1369.23
Moderately fast ramp0.2567.0072.76
Default schedule (GPS)0.5071.5577.24
Slower ramp0.7570.8675.96
Ramp throughout1.0070.6275.00

The default schedule achieves the highest mAP and Rank-1 among the tested schedules. It improves on full strength from the start by 9.62 and 8.33 percentage points, respectively. The results support increasing the synthesis strength gradually after the initial ReID representation has been learned.

Dear Reviewer,

Thank you for taking the time to review our paper and for your constructive feedback. We believe our responses address the concerns you raised.

With fewer than two days remaining in the discussion period, we wanted to check whether you have any further questions or concerns. If there is anything you would like us to clarify, please feel free to let us know. We would be happy to respond promptly before the discussion closes.

Best regards, The Authors

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

This paper addresses Aerial-Ground Person Re-Identification (AG-ReID), where matching the same identity across drone and ground views suffers from drastic local appearance and visibility changes. The authors propose GPS, a training-time hard-positive synthesis framework operating in the CLIP patch-token feature space rather than pixel space. GPS consists of three modules: (1) GCPS, which reorganizes patch tokens via compression, merging, and vanishing under a fixed spatial coordinate prior to simulate viewpoint-induced evidence redistribution; (2) HPPS, which selects the hardest identity-consistent synthetic candidate within a prescribed difficulty band relative to a real same-identity cross-view observation; and (3) DPEF, which fuses locally supported correspondence between the original and selected synthetic features into identity learning while preserving unsupported original evidence. Experiments on AG-ReID v1/v2, LAGPeR, and CARGO show consistent gains over Plain CLIP, with additional ablations across modules, hyperparameters, backbones, and ReID frameworks.

Critical strengths
  1. Clear and well-motivated framing. Recasting aerial-ground viewpoint change as a redistribution of local identity evidence (spatial contraction, unequal cue contribution, cue vanishing) is conceptually clean, and the three operations (compression/merging/vanishing) map intuitively to real geometric effects observed in AG-ReID.
  2. Meaningful departure from generic augmentation. Unlike tranditional methods such as Random Erasing, which operate at the pixel level without structural or difficulty control, GPS constructs positives in feature space with identity-consistency constraints and a difficulty interval anchored to real cross-view pairs. This is a non-trivial and practically useful idea.
  3. Reasonably comprehensive empirical evaluation. Four benchmarks, multiple cross-view protocols, full component ablations, parameter sensitivity sweeps, cross-backbone and cross-ReID-framework generalizability checks are all provided.
  4. Zero inference overhead. All synthesis and fusion modules are training-only, at inference the model reverts to a single CLIP encoder with cosine ranking, which is attractive for deployment.
Critical weaknesses
  1. "Geometry-guided" is under-delivered. The method uses only a fixed 2D coordinate prior and a Gaussian anisotropic neighborhood weight. There is no explicit perspective projection, scale modeling, or camera-height-induced mapping between aerial and ground planes. Compression, merging, and vanishing are token-space operations guided by coordinates, not geometric projections.
  2. Does not uniformly beat existing AG-ReID SOTA.
  3. No training-efficiency analysis. Generating multiple GCPS candidates, searching within them via HPPS, and running DPEF's dual-path fusion all add training-time cost.
  4. Orthogonality to existing AG-ReID architectures is under-justified. Table 3 varies the pretrained backbone, but these differ mainly in pretraining rather than in AG-ReID-specific structural design. The paper does not demonstrate that GPS can be stacked on top of existing AG-ReID-dedicated modules and still yield gains.
  5. No failure-case analysis. Figure 4 presents successful retrievals and favorable margin distributions, but the paper never discusses when GPS fails.
  6. Mixing ratio between synthetic and real cross-view positives is unreported. How often are GCPS-synthesized positives sampled versus real same-identity cross-view pairs? If synthetic positives dominate, is there a risk of overfitting to synthetic artifacts?
Questions for authors
  1. What are the specific default values of the HPPS difficulty bounds aa and bb? Do they need to be retuned per dataset, and what is the sensitivity range shown in Figure 3?
  2. Compression and merging both reduce the spatial extent of local evidence. What distinct geometric effects do they capture, and why is both necessary rather than one or the other?
  3. On AG-ReID v2 A→W, GPS underperforms VDT. What is the authors' hypothesis for this gap?
  4. How is "locally supported correspondence" in DPEF determined? What exactly constitutes "unsupported original evidence"?
  5. What is the actual training throughput (images/sec) and GPU memory overhead of GPS compared to Plain CLIP on the same hardware?
5 comments

We sincerely thank you for your thorough review and suggestions.

Q1: "Geometry-guided" is under-delivered. The method uses only a fixed 2D coordinate prior and a Gaussian anisotropic neighborhood weight. There is no explicit perspective projection, scale modeling, or camera-height-induced mapping between aerial and ground planes. Compression, merging, and vanishing are token-space operations guided by coordinates, not geometric projections.

A1:

We consider the 2D coordinate transformation and the compression, merging, and vanishing (CMV) operations motivated by aerial–ground viewpoint changes to be forms of geometric guidance. Common data and feature augmentation methods often rely on random edits. GPS uses an explicit 2D coordinate transformation to guide feature augmentation. This transformation itself provides geometric guidance. There are three additional reasons for this design. First, person re-identification typically follows pedestrian detection, so the person's position and body layout are generally expected to be stable within the detected crop. We observed this pattern across several real-capture benchmarks and manually audited a stratified random sample of 360 training crops. A crop was considered compatible if the principal body axis was within 30° of vertical and the horizontal and vertical offsets of the center of the body's extent from the crop center were at most 20% of the crop width and height, respectively. The audit results are shown below.

Table A: Manual audit of crop-layout compatibility across the four benchmarks.

DatasetCompatible crops
AG-ReID v189.6%
AG-ReID v287.4%
LAGPeR95.8%
CARGO66.9%

Second, explicit viewpoint projection and scale modeling often require more detailed observations, including camera intrinsics, viewing angle, camera pose, or height. Mainstream ReID benchmarks generally lack these calibration details, particularly precise camera heights. GPS is designed for direct use in practical settings without additional annotations or recalibration.

Third, the fixed coordinate grid provides a soft spatial reference for compression, neighborhood interactions, and positional support. A given grid position can represent cues from different body parts, including the head, torso, and legs. For example, if a shifted detection box places a shoulder cue on one side of the crop, that cue can still match a donor patch based on its own token response.


Q2: (a) Does not uniformly beat existing AG-ReID SOTA.

(b) On AG-ReID v2 A→W, GPS underperforms VDT. What is the authors' hypothesis for this gap?

A2:

GPS outperforms VDT on AG-ReID v2 A→W by 6.08 percentage points in mAP and 3.35 in Rank-1. GPS leads 23 of 26 protocol–metric comparisons and improves both metrics over Plain CLIP on all 13 protocols (§4.3, Table 1).

We think you may be referring to the Rank-1 result on CARGO A→A. On CARGO A→A, Table 1(d) shows higher mAP and lower Rank-1 for GPS compared with VDT. We additionally counted the valid queries and found only 40 valid aerial queries after filtering. Just one additional correct first-ranked match raises Rank-1 by 2.50 percentage points. With so few valid queries, close Rank-1 scores on this protocol provide limited resolution for distinguishing the methods' retrieval performance. GPS also outperforms the Plain CLIP baseline on both metrics in this protocol. GPS makes only one more Rank-1 error than VDT on this protocol. On all other metrics, GPS substantially outperforms the best existing methods, indicating strong overall performance and good generalization.

Q3: (a) No training-efficiency analysis. Generating multiple GCPS candidates, searching within them via HPPS, and running DPEF's dual-path fusion all add training-time cost.

(b) What is the actual training throughput (images/sec) and GPU memory overhead of GPS compared to Plain CLIP on the same hardware?

A3:

We further measured the training and inference costs of the compared methods. Compared with Plain CLIP, GPS slightly increases training time while leaving peak training GPU memory nearly unchanged. Inference uses the same descriptor extraction pipeline as Plain CLIP at comparable cost. Compared with mainstream ReID methods such as GSAlign and ViSA, GPS uses fewer parameters and has substantially lower training and inference costs, with a particularly clear advantage in inference throughput.

Table 5 in Appendix C.4 reports trainable parameters and training time per epoch. We additionally measured peak GPU memory and single-image inference latency. All methods were benchmarked with the same hardware configuration and optimization settings.

Table B: Training and inference costs and CARGO ALL retrieval performance of the compared ReID methods.

MethodParams (M)Train s/epochTrain peak (GiB)Infer ms/imgInfer peak (GiB)Equiv. img/sCARGO ALL mAP (%)CARGO ALL Rank-1 (%)
Plain CLIP89.34120.814.844.550.57219.7863.4969.55
VDT87.6760.334.804.640.56215.5255.2064.10
SeCap134.76119.719.957.770.82128.7060.1968.59
HiHR136.37140.178.7810.920.7691.5867.5274.36
GSAlign111.67168.155.9913.110.7076.2857.9565.06
ViSA275.62275.1911.5939.911.3625.0665.4670.51
GPS89.93164.574.884.540.57220.2671.5577.24

Q4: Orthogonality to existing AG-ReID architectures is under-justified. Table 3 varies the pretrained backbone, but these differ mainly in pretraining rather than in AG-ReID-specific structural design. The paper does not demonstrate that GPS can be stacked on top of existing AG-ReID-dedicated modules and still yield gains.

A4:

We evaluated GPS's compatibility with several mainstream ReID frameworks. Adding GPS yielded clear gains in both retrieval metrics across all evaluated frameworks, supporting its use as an orthogonal training component that can further improve existing methods.

Section 4.8 and Table 3 already report the VDT and SeCap integrations. In addition, we evaluated the effect of integrating GPS into ViSA, HiHR, and GSAlign on CARGO.

Table C: GPS integration with dedicated AG-ReID frameworks on CARGO ALL.

FrameworkBase mAP / Rank-1 (%)Gain (pp)With GPS (%)
VDT55.20 / 64.10+4.04 / +2.6159.24 / 66.71
SeCap60.19 / 68.59+8.15 / +3.5368.34 / 72.12
ViSA65.46 / 70.51+3.44 / +6.4168.90 / 76.92
HiHR67.52 / 74.36+3.59 / +1.9271.11 / 76.28
GSAlign57.95 / 65.06+10.87 / +11.5468.82 / 76.60

Q5: No failure-case analysis. Figure 4 presents successful retrievals and favorable margin distributions, but the paper never discusses when GPS fails.

A5:

Our failure-case review identified two recurring conditions. Very low illumination can obscure clothing textures in the source or donor. Multiple people in the same crop, particularly with similar clothing, can make it harder to associate donor patches with the target person. These conditions affect the local evidence available for donor-guided synthesis. Even in these difficult cases, GPS can still retrieve more correct candidates than other methods.

These issues can also be mitigated at the dataset level. Better-exposed captures can preserve clothing detail, and tighter person crops can reduce interference from other people in the frame. You can find this additional failure-case analysis in §4.9 and Figure R1 of the paper.


Q6: Mixing ratio between synthetic and real cross-view positives is unreported. How often are GCPS-synthesized positives sampled versus real same-identity cross-view pairs? If synthetic positives dominate, is there a risk of overfitting to synthetic artifacts?

A6:

Our method constructs hard positives only during training to align complex aerial–ground features. At test time, inference and evaluation use only real samples, without any synthetic samples or features. Therefore, there is no risk of overfitting to synthetic samples. The substantial performance gains also show that using synthetic samples in training bridges the gap between training and real test data (§4.3, Table 1).

We report a sensitivity analysis of the proportion of eligible sources entering GCPS on the CARGO ALL and A→G protocols.

Table D: Sensitivity to the proportion of eligible sources entering GCPS on CARGO ALL and A→G.

Eligible sources entering GCPSALLA→G
0% (GCPS disabled)64.23 / 69.2365.12 / 64.89
25%67.28 / 73.4068.24 / 69.15
50%69.52 / 75.6471.26 / 72.34
75% (paper default)71.55 / 77.2473.35 / 74.47
100%71.09 / 76.9272.84 / 73.40

We also report CARGO training statistics to quantify how often sources receive synthetic supervision. During active synthesis, every source had an observed cross-view reference, and about three-fifths also received a synthetic-feature update.

Table E: Coverage of synthetic-feature updates during CARGO training.

CARGO training statisticValue
Original inputs per batch64
Sources with an observed same-ID cross-view reference100%
Eligible sources entering GCPS75%
Candidates per selected source / maximum retained5 / 1
Mean HPPS acceptance among participating sources79.88%
Estimated update coverage75% × 79.88% = 59.91% (about 38 of 64 inputs)

Q7: What are the specific default values of the HPPS difficulty bounds [δmin⁡\delta_{\min}] and [δmax⁡\delta_{\max}]? Do they need to be retuned per dataset, and what is the sensitivity range shown in Figure 3?

A7: HPPS difficulty bounds

We use the same default bounds across all benchmarks (§3.3 and Appendix B.2). Figure 3(b) and Appendix C.7 report the CARGO sweep, which varies the lower bound at a fixed interval width.

SettingValue
Default lower / upper bounds0.004 / 0.007
Fixed interval width0.003
Lower-bound sweep0 to 0.014 in steps of 0.002

Q8: Compression and merging both reduce the spatial extent of local evidence. What distinct geometric effects do they capture, and why is both necessary rather than one or the other?

A8: Compression and merging

Compression changes the spatial extent of a cue, while merging changes the relative contributions of cues within a local observation. The two operations capture distinct aerial–ground changes in different scenarios, so both are needed. Consider a person wearing a gray jacket with a chest logo and a black shoulder strap worn diagonally. In a horizontal ground-level view, the chest logo and jacket hem are several patch rows apart. From an oblique aerial view, the distance may shrink so that the cues fit within fewer patch rows, with the logo still above the hem. This is the change compression models. The shoulder strap and gray fabric may be distinct in adjacent regions before the viewpoint changes. In another view, they can fall into the same small image region, where the strap remains prominent and the fabric texture becomes blurred. Merging uses local support to model this unequal mixing. Strongly supported tokens retain more of their own features, while weakly supported tokens absorb more information from their neighbors.

The paper's results on CARGO A→G show that combining the two operations achieves higher mAP and Rank-1 than using either alone (§4.4, Table 2(a), and Appendix C.3).

Table F: Compression and merging ablation on CARGO A→G.

OperatorsmAP (%)Rank-1 (%)
Compression68.8070.21
Merging69.5669.15
Compression + merging70.6572.34

Q9: How is "locally supported correspondence" in DPEF determined? What exactly constitutes "unsupported original evidence"?

A9: Local correspondence and original-only evidence

DPEF softly aligns synthetic patches with original patches using feature similarity, coordinate distance, and a positional prior (Eq. 15). For each original patch, DPEF takes the highest matching score and normalizes it by the geometric support ceiling at that position (Eqs. 16–17). At relative support of at least 0.82, the position is treated as shared evidence (Eq. 6), and the synthetic feature can provide a weighted update to the original feature. If relative support falls below 0.82 and the normalized original-patch response is at least as strong as the aligned synthetic-patch response, the position is treated as original-only evidence. Fusion preserves the original cue with a small adjustment.

For example, a person's black shoulder strap and chest logo may both be clear in the original image. After synthesis, the strap shifts slightly, but a sufficiently similar strap feature is still present nearby and can serve as shared evidence. The chest logo may become blurred, leaving the synthetic feature below the correspondence threshold and with a weaker response than the original. DPEF therefore preserves the logo from the original image. Here, "lack of support" means that the synthetic result lacks a sufficiently reliable correspondence for the logo, which remains discriminative in the original image.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.