acceptodds
Under review as a conference paper at ICLR 2027

Text-Anchored Feature Deformation for Aerial–Ground Person Re-Identification

Abstract

Aerial-ground person re-identification requires matching identities across substantial changes in viewpoint and visual appearance. A central challenge is to establish cross-view correspondence while retaining the view-specific cues needed for identity discrimination. Current methods typically utilize direct feature alignment strategies, which do not explicitly model how visual evidence should be reorganized across viewpoints. Therefore, we propose Text-Anchored Feature Deformation (TAFD), which formulates cross-view matching through language-conditioned directional transformations in feature space. A Semantic View Anchoring (SVA) module learns aerial and ground semantic anchors using shared and view-private text prompts encoded by a frozen text encoder. Conditioned on the difference between these anchors, a Directional Feature Deformation (DFD) module employs two independent operators, one for aerial-to-ground transformation and the other for ground-to-aerial transformation, to predict spatial offsets, resample multi-level patch features, and aggregate target-view evidence. To preserve identity during transformation, a Target-Space Identity Consistency (TIC) objective supervises transformed descriptors using target-view classifiers and batch identity prototypes constructed from stop-gradient target features, while native features retain view-specific classification and metric supervision. Experiments on CARGO and AGReIDv1 demonstrate the effectiveness of TAFD for aerial-ground person retrieval. On CARGO, TAFD achieves 72.59% mAP and 76.60% Rank-1 for aerial-to-ground retrieval. On AGReIDv1, the same architecture achieves average bidirectional performance of 80.79% mAP and 87.51% Rank-1.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Person Re-Identification, View-aware Learning, Aerial–Ground Retrieval

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SDd2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes Text-Anchored Feature Deformation (TAFD) for aerial–ground person re-identification. It argues that forcing both viewpoints into a shared representation can weaken view-specific identity cues. TAFD therefore retains a native descriptor for each image while learning an additional descriptor for the opposite view. The framework consists of three components. Semantic View Anchoring (SVA) uses learnable prompts and a frozen CLIP text encoder to obtain aerial and ground anchors. Directional Feature Deformation (DFD) uses these anchors to guide spatial resampling and attention-based aggregation of multi-level visual features. Target-Space Identity Consistency (TIC) supervises transformed descriptors using target-view classifiers, instances, and identity prototypes. Experiments on CARGO and AGReIDv1 report competitive results, including 72.59% mAP on CARGO A→G and 80.79% average mAP on AGReIDv1. Although the explicit target-view supervision and preservation of native features offer a useful formulation, several issues remain unresolved. The baseline used in component ablations requires clearer justification and consistent training conditions. The semantic contribution of text anchors is not sufficiently distinguished from ordinary learnable conditioning, while the role of view-specific LoRA remains underexplored. A comparison with the closely related CVAF method would further strengthen the evaluation.

Critical strengths
  1. The explicit target-view identity supervision for cross-view feature transformation. Rather than relying solely on geometric correction or generic feature alignment, the proposed Target-Space Identity Consistency (TIC) supervises transformed descriptors using the destination-view classifier, native instances, and identity prototypes. This provides a concrete discriminative objective for each transformation direction and directly addresses the ambiguity of where deformed features should be mapped. The formulation establishes a meaningful connection between feature deformation and identity preservation, making it a valuable contribution to aerial–ground ReID.
  2. The design of separate native and transformed representations. By maintaining view-specific identity features while learning independent aerial-to-ground and ground-to-aerial transformations, TAFD offers a unified framework for both cross-view and within-view retrieval. The use of multi-level feature deformation further allows the model to capture spatial variations across viewpoints. The experimental results are also competitive, achieving 72.59% mAP and 76.60% Rank-1 on CARGO aerial-to-ground retrieval, and 80.79% average mAP on AGReIDv1. These results demonstrate the practical potential of the proposed formulation and provide meaningful empirical support for its effectiveness.
Critical weaknesses
  1. Table 3 describes its baseline as CLIP ViT-B/16 trained under the same framework, yet the reported results exactly match those of the ImageNet-pretrained ViT+Explainable model in the original AG-ReID paper. The paper does not establish whether the baseline and proposed variants use identical pretrained weights, LoRA configurations, and training protocols. This directly undermines the attribution of the reported 7.08-point mAP improvement to DFD.
  2. SVA generates two global, image-independent anchors, while the transformation directions are already determined by separate deformation networks and target-view supervision. The proposed conditioning could therefore be implemented using ordinary learnable vectors. Without comparing against such alternatives, the experiments do not establish whether CLIP's textual semantics contribute beyond additional learnable parameters and view-specific conditioning.
  3. DFD combines learnable spatial offsets, multi-level feature aggregation, and attention-based pooling. However, the ablations do not isolate these operations. In particular, a zero-offset variant retaining the same aggregation architecture would determine whether spatial deformation provides meaningful improvements beyond enhanced feature extraction and pooling. This distinction is central to the paper's motivation.
  4. Including the comparison with CVAF [DOI:10.1145/3785482] would provide a more relevant assessment of TAFD against recent approaches using vision-language representations.
Questions for authors
  1. In Table 3, was the baseline independently trained using CLIP ViT-B/16 under the same pretraining, LoRA configuration, and training protocol as the proposed variants, or were its results taken from the original AG-ReID paper?
  2. How does SVA compare with directly learnable view embeddings under identical training settings? What evidence shows that the improvement comes from CLIP's textual semantics rather than view-specific learnable conditioning?
  3. How does DFD perform when spatial offsets are fixed to zero while retaining the same multi-level features and attention aggregation? How much of the reported improvement can be attributed specifically to spatial deformation?
  4. How does TAFD compare with CVAF (DOI: 10.1145/3785482) on AGReIDv1? Can the authors include its reported results and clarify any differences in evaluation protocols and training settings?
2 comments

Q1: The baseline of the method.

Thanks a lot for identifying this inconsistency. The baseline results in the submitted Table 3 in this paper were taken from the original AG-ReID paper, not independently obtained using CLIP ViT-B/16. They correspond to its ImageNet-pretrained ViT+Explainable model. We initially used this method as our reference baseline, but subsequently failed to update the baseline. We apologize for this reporting error. We have now independently trained a conventional CLIP ViT-B/16 baseline using the same CLIP pretrained weights, view-specific LoRA configuration, and common training protocol as our method. This baseline uses identity classifiers shared across aerial and ground images, with ID and triplet losses, while excluding SVA, DFD, and TIC. We have replaced the original baseline entry in the revised manuscript. The corrected results are provided below.

SettingA→G mAPA→G Rank-1G→A mAPG→A Rank-1Avg. mAPAvg. Rank-1
Baseline74.4682.6377.0084.2075.7383.42
+ DFD79.4785.6280.6987.9480.0886.78
+ DFD + TIC79.1486.4680.9788.4680.0687.46
+ DFD + TIC + SVA79.5586.4682.0388.5780.7987.51

Q2: Comparison between SVA and Learned Vectors.

Thank you for your constructive comments. We clarify that SVA learns identity-supervised, task-adaptive view anchors. Although the CLIP text encoder is frozen, the shared and view-private prompt tokens are trainable and receive gradients from identity-supervised retrieval objectives. The resulting anchors therefore adapt to the re-identification task rather than remaining fixed text embeddings. We agree that trainability and identity supervision alone do not distinguish SVA from ordinary learnable vectors. SVA’s distinction lies in its parameterization: the two anchors are generated through shared/private prompts and the frozen CLIP text encoder, rather than optimized directly as independent vectors.

To evaluate this distinction, we compared SVA with two directly learnable vectors in the Table 2 below. As shown in Table 2, SVA improves mAP and Rank-1 in both retrieval directions compared to learnable vectors. The improvements are 0.48% mAP and 1.18% Rank-1 for A→G, and 0.31% mAP and 0.80% Rank-1 for G→A. These results support the benefit of the SVA over learnable vectors .

Table 1. Comparison between SVA and Learnable Vectors on AGReIDv1 under the same training protocol.

SettingA→G mAPA→G Rank-1G→A mAPG→A Rank-1Avg. mAPAvg. Rank-1
Learnable Vectors79.0785.2881.7287.7780.4086.53
SVA79.5586.4682.0388.5780.7987.51

Q3: DFD without spatial offsets.

Thank you for your suggestion. To isolate the contribution of spatial deformation, we conduct an additional experiment where all spatial offsets are fixed to zero, while retaining the same multi-level features, attention aggregation, and training objectives. In this setting, DFD aggregates features from their original spatial locations without learning spatial rearrangement. As shown in Table 2, introducing learnable spatial offsets improves the average mAP from 76.11% to 80.08% and Rank-1 from 84.35% to 86.78%. Specifically, spatial deformation contributes improvements of 4.74% mAP and 2.05% Rank-1 for A→G, and 3.21% mAP and 2.80% Rank-1 for G→A. These results indicate that multi-level feature aggregation alone provides useful representations, but spatial deformation brings substantial additional gains by allowing the model to adaptively reorganize local features across different viewpoints.

Table 2. Contribution of spatial deformation on AGReIDv1.

SettingA→G mAPA→G Rank-1G→A mAPG→A Rank-1Avg. mAPAvg. Rank-1
DFD w/o spatial offsets74.7383.5777.4885.1476.1184.35
DFD79.4785.6280.6987.9480.0886.78

Q4: Comparison between SVA and CVAF.

A: Thank you for bringing CVAF [1] to our attention. We apologize for overlooking this relevant work. After carefully reading the paper, we found that CVAF reports 79.45% mAP / 87.23% Rank-1 under A→G and 82.59% mAP / 90.12% Rank-1 under G→A on AGReIDv1. For comparison, LATex [2] reports 77.67% / 85.26% and 81.15% / 89.40%, while TAFD achieves 79.55% / 86.46% and 82.03% / 88.57%, respectively. The average mAP is therefore 79.41% for LATex†, 80.79% for TAFD, and 81.02% for CVAF. Although CVAF obtains higher Rank-1, TAFD achieves comparable average mAP and slightly higher A→G mAP.

The baseline comparison provides further context. Our baseline in Table 3 are 74.46% and 77.00% mAP under A→G and G→A, respectively, averaging 75.73%. LATex reports 72.36% and 74.78% mAP for its AIE-only CLIP baseline, averaging 73.57%. Both are largely below CVAF’s CLIP-ReID baseline, which achieves 78.84% and 81.00% mAP, averaging 79.92%. Within their respective settings, TAFD improves average mAP from 75.73% to 80.79%, whereas CVAF improves it from 79.92% to 81.02%. Thus, TAFD achieves competitive final performance from a lower baseline, with a substantial improvement attributable to its proposed components within our experimental setting. We could not locate a publicly available implementation of CVAF, so we have not been able to verify the remaining implementation differences or establish a matched reproduction. We have include CVAF in the related work in the revised manuscript, together with a clarification of these differences.

[1] CVAF: A CLIP-Based View-Consistent Alignment Framework for Aerial-Ground Person Re-Identification. ACM TOMM 2026.

[2] LATex: Leveraging Attribute-based Text Knowledge for Aerial-Ground Person Re-Identification. Arxiv 2025.

Finally, we hope our response has adequately addressed your concerns. If you have any further suggestions, please feel free to discuss them with us. We will provide the corresponding results as soon as possible.

Reviewer_V8vBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TAFD for aerial-ground person re-identification. It retains each image’s native descriptor and learns another descriptor for the opposite view. Learnable text anchors guide directional feature deformation, while target-view classifiers and identity prototypes supervise the converted features. Results are reported on CARGO and AGReIDv1.

Critical strengths
  1. Separating native and converted representations is a sensible design. Target-view identity supervision gives each conversion an explicit destination while retaining the original descriptor.
  2. The reported CARGO A→G gains are encouraging: 3.59 points in mAP and 5.32 points in Rank-1 over ViSA. The mAP improvements on AGReIDv1 provide further support.
Critical weaknesses

1.The methodological advance needs a stronger justification. GSAlign already performs spatial deformation, ViSA preserves view-specific cues, and IDEA combines learned offsets with sampled-feature attention. The remaining novelty lies mainly in retaining native features alongside destination-view conversion. In rebuttal, the authors should explain the distinct insight this formulation provides about cross-view matching and identify which evidence supports that insight.

2.The language-specific contribution remains unclear. Both anchors are global, and the two directions already have separate operators. Once trained, the anchor-difference contribution in Eq. 4 can be absorbed into a direction-specific bias. What does the frozen text encoder contribute beyond two learned conditioning/query vectors? The rebuttal needs a concrete justification, supported by a focused comparison or convincing analysis, since text anchoring is central to the paper’s claimed contribution.

3.Table 3 leaves a basic ambiguity in the “+DFD” setting. DFD requires anchors in Eqs. 4 and 6, while Eq. 11 places transformed-feature supervision under TIC. What anchors and training objectives remain when SVA and TIC are absent? Please reconcile this setting with the method and explain precisely what supervision TIC adds. This matters because “+DFD” accounts for most of the reported improvement.

Questions for authors

I would have preferred a borderline rating. Since this option is unavailable, I lean toward weak accept, given the useful representation design and encouraging results. However, I would lower my rating if the rebuttal does not convincingly address the novelty, the specific role of text anchors, and the ablation definitions. The authors will see the Specific Points of Feedback for Rebuttal in Weaknesses.

3 comments

Q1: The methodological advance needs a stronger justification.

A: Thank you for your feedback. Our main concern is that view-specific features may lose useful identity information during cross-view learning. When shared and view-specific features are learned together, the network may mainly rely on shared cues that are easier to match across views. The view-specific features can then contribute little to identity recognition, even though the corresponding branches are retained. This is the feature degeneration risk that motivates our design.

GSAlign [1] addresses geometric and semantic misalignment, but does not explicitly require a corrected aerial feature to match the same identity in the ground-view feature space, or vice versa. Its reported performance drop in aerial–aerial retrieval also highlights the need to examine whether cross-view alignment comes at the expense of within-view discrimination. ViSA [2] explicitly preserves view-specific cues through view classification and orthogonality constraints. However, distinguishing viewpoints does not necessarily mean distinguishing identities. Its identity losses supervise the global and fused local features, leaving open the possibility that the fused representation mainly relies on shared cues. IDEA [3] uses learned sampling positions and attention to aggregate multimodal features, rather than supervising a source-view feature to represent the same identity in the opposite view.

To address this issue, TAFD separately supervises native features and cross-view transformations. Native aerial and ground features retain their own losses, so they are directly trained to distinguish identities within each view. The transformed features are supervised by the target-view classifier, instances, and identity prototypes. Target-side gradients are detached along these consistency paths, so the target references do not receive gradients from the conversion loss. Thus, we do not simply concatenate two branches: with explict transformation and constrains, we train one to preserve native identity information and the other to match identity information from the opposite view.

Q2: The contribution language-specific anchor.

Thank you for pointing out this issue. We agree that, after training, the anchor-difference term in Eq. (4) can be absorbed into a direction-specific bias. The role of the text encoder concerns how these conditioning vectors are learned. In SVA, shared and view-private prompt tokens are optimized through the same frozen pretrained CLIP text encoder. Although the encoder is frozen, the anchors change as the prompts are updated. This provides a pretrained mapping for learning the anchors used in both offset prediction and feature aggregation. Following your suggestion, we replace the text anchors with two directly learnable vectors in both Eqs. (4) and (6), retaining the same deformation operators, attention modules, TIC objectives, and training settings. As shown in Table 1, SVA improves mAP and Rank-1 in both retrieval directions compared to learnable vectors. The improvements are 0.48% mAP and 1.18% Rank-1 for A→G, and 0.31% mAP and 0.80% Rank-1 for G→A. These results support the benefit of the SVA over .

Table 1. Comparison between SVA and Learnable Vectors on AGReIDv1 under the same training protocol.

SettingA→G mAPA→G Rank-1G→A mAPG→A Rank-1Avg. mAPAvg. Rank-1
Learnable Vectors79.0785.2881.7287.7780.4086.53
SVA79.5586.4682.0388.5780.7987.51

Q3: The explanation of '+DFD' setting in Table 3 of the paper.

Thank you for pointing out this ambiguity. '+DFD' is an anchor-free deformation variant trained with standard ReID supervision. In this setting, SVA is removed by setting both text anchors to zero, bypassing the prompt learner and text encoder. Consequently, the anchor-difference input in Eq. (4) and the raw attention-query input in Eq. (6) are zero. The two deformation operators still predict image-dependent offsets from visual patch features and perform bilinear sampling and feature aggregation. Since the attention query projection includes a trainable bias, zero text input does not force attention to remain uniform. When TIC is absent, the deformed features are trained using general losses. Specifically, the deformed features from both views are classified by a shared identity classifier, and the triplet loss is computed over these features. This procssing is somewhat like GSAlign [1], which also consider the transformation under general constrains. The source-side references are detached along these paths, while native features and their classifiers retain their own training objectives. Thus, the deformation operators remain supervised without SVA or TIC.

TIC introduces a specific target-view reference for each transformation. For aerial-to-ground transformation, the deformed aerial feature is supervised by the ground-view identity classifier and compared with native ground features through the directional triplet loss. The positive-prototype cosine and prototype contrastive losses further constrain it using ground-view identity prototypes. The reverse direction is supervised analogously. The target-side classifier, native features, and prototypes are detached along these consistency paths, so the corresponding gradients act on the transformed source representation. Thus, while the basic setting jointly aligns the two sets of deformed features, TIC explicitly requires each transformed feature to match the native identity representation of its destination view.

[1] GSAlign: Geometric and Semantic Alignment Network for Aerial-Ground Person Re-Identification. Neurips 2025.

[2] View-Aware Semantic Alignment for Aerial-Ground Person Re-Identification. CVPR 2026.

[3] IDEA: Inverted Text with Cooperative Deformable Aggregation for Multi-Modal Object Re-Identification. CVPR 2025.

Finally, we hope our response has adequately addressed your concerns. If you have any further suggestions, please feel free to discuss them with us. We will provide the corresponding results as soon as possible.

I thank the authors for providing detailed supplementary experiments and additional clarifications. These clarifications have addressed my concerns. I maintain my score of 3.

Reviewer_Pw6pOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TAFD for person re-identification across aerial and ground camera views. It combines native-view representations with separate directional transformations guided by learned text anchors. Target-space identity supervision encourages transformed features to retain identity information in the destination view. Experiments on CARGO and AGReIDv1 show promising cross-view retrieval performance, although broader validation is needed.

Critical strengths

The paper addresses a relevant aerial-ground ReID challenge by retaining native features and learning independent directional transformations to accommodate asymmetric viewpoint differences. Target-view classifiers and identity prototypes provide explicit identity-preserving supervision. TAFD reports strong cross-view performance, achieving 72.59% mAP and 76.60% Rank-1 on CARGO and the highest reported average mAP on AGReIDv1. Ablations on component contributions, transformation directions, loss weights, and prompt sizes provide useful initial support for the design choices.

Critical weaknesses
  1. TAFD requires each image to be identified as aerial or ground to select the appropriate adapter, deformation direction, and descriptor arrangement. This assumption is reasonable when camera metadata is available, but performance with unavailable or incorrect view labels is not investigated, leaving its applicability in such settings unclear.
  2. The interaction among SVA, DFD, and TIC is difficult to follow as a complete training pipeline. In particular, Figure 2 does not illustrate the TIC supervision pathway, including target-view classifiers, identity prototypes, and gradient flow. A complete training diagram and step-by-step algorithm would improve clarity and reproducibility.
  3. Table 1 omits ground-to-aerial results despite this direction being listed in Section 4.1. Furthermore, within-view improvements are not consistent across metrics: aerial-to-aerial Rank-1 is 80.00% versus 82.50% for VDT/ViSA, and ground-to-ground mAP is 83.57% versus 83.90% for ViSA. Although TAFD achieves the best aerial-to-aerial mAP, the evidence for uniformly stronger within-view retrieval is limited.
  4. The generalisability of the proposed method remains insufficiently established, as evaluation is restricted to AGReIDv1 and the synthetic CARGO dataset. Without evaluation on additional real-world benchmarks, such as AG-ReIDv2, G2APS-ReID, CP2108, and LAGPeR, it is unclear whether the reported gains extend across different camera networks and acquisition conditions. Cross-dataset testing without retraining would further help establish the method’s ability to generalise to unseen environments.
  5. Although preserving within-view identity discrimination is a central motivation of TAFD, the paper reports aerial-to-aerial and ground-to-ground results only on the synthetic CARGO dataset. No within-view results are provided on any real-world benchmark. Consequently, the claim that the method preserves view-specific discrimination remains insufficiently validated under real-world conditions and should be assessed on a suitable dataset with multiple cameras per view.
  6. As shown in Table 3, adding TIC to DFD yields only limited improvement, with essentially unchanged average mAP and a modest increase in Rank-1. Adding SVA provides a further small gain, with an almost negligible improvement in Rank-1. Since DFD accounts for most of the overall performance increase, the necessity and additional value of TIC and SVA require stronger experimental justification.
Questions for authors
  1. How are aerial/ground labels obtained at inference through camera metadata or a separate prediction process? How would the framework operate when this information is unavailable or unreliable?
  2. Could the authors provide a complete training and inference diagram or algorithm showing how SVA, DFD, and TIC interact? In particular, please clarify the feature flow, target-view supervision, and gradient propagation through the different components.
  3. Section 4.1 lists ground-to-aerial retrieval among the evaluation protocols, but Table 1 omits these results. Could the authors report this direction and discuss why the strong cross-view improvements do not translate consistently across within-view metrics?
  4. Why is the evaluation restricted to AGReIDv1 and synthetic CARGO? Could the authors provide evidence from additional real-world benchmarks, such as AG-ReIDv2 and LAGPeR, and clarify whether the method generalises across datasets without retraining?
  5. What experimental evidence supports the claim that TAFD preserves within-view discrimination in real-world settings? Could you evaluate aerial-to-aerial and ground-to-ground performance on a suitable real-world benchmark with multiple cameras per view?
  6. In Table 3, adding TIC changes average mAP from 80.08% to 80.06%, while improving Rank-1 by 0.68 percentage points. Adding SVA subsequently improves mAP by 0.73 points but Rank-1 by only 0.05 points. How do the authors justify the additional components relative to these modest gains, and are the improvements consistent across repeated runs?
3 comments

Q1: Where the aerial and ground labels come?

A1: Thank you for your detailed feedback. We would like to clarify that TAFD uses the same aerial/ground information as existing aerial-ground ReID methods. On CARGO and AGReIDv1, we use the camera/view information provided by the datasets, without a separate view-prediction model. In the intended aerial-ground camera network, this information is known from whether an image originates from a UAV or a ground-camera stream. TAFD requires only this binary platform label to select the adapter, deformation direction, and descriptor arrangement in Eq. (12); detailed camera calibration or viewing angles are unnecessary. The availability of this information follows the established aerial-ground ReID setting. For example, VDT [1] explicitly states that aerial/ground labels can be obtained from known camera labels. CARGO [2] also provides a direct correspondence, with cameras 1–5 belonging to the aerial view and cameras 6–13 to the ground view. Therefore, TAFD follows the existing setting in which the acquisition source is known. Handling unavailable or unreliable view information is an interesting research question. However, our work follows the established aerial-ground ReID setting, where the acquisition source is known.

Q2: How do SVA, DFD, and TIC interact during training and inference?

Thank you for question. We provide a Algorithm 1 shows the processing of our model. Let ss be the observed view, tt the opposite view, and HsH_s the native identity head for view ss.

Algorithm 1. TAFD training and inference

SVA(prompt tokens)
    return FrozenCLIPText(shared and view-private prompt tokens) as a_A, a_G

FORWARD(x, s, a_A, a_G)
    t ← opposite(s)
    z, C, P ← CLIPVision(x, input-view bias and LoRA for s)
        # z is the native descriptor. C and P are projected CLS and patch tokens
        # from layers 5, 8, and 12, obtained in the same visual pass.
    Δ ← 2 · tanh(OffsetMLP[s→t](Concat(LN(P), Broadcast(LN(a_t − a_s)))))
    P_warp ← BilinearSample(P, patch_grid + Δ) at each selected layer  # patch units
    c ← Attn[s→t](query=a_t, key=C, value=C)
    p ← Attn[s→t](query=a_t, key=P_warp, value=P_warp)
    r ← 32 if s = A, otherwise 16
    z_conv ← LayerNorm[s→t](r · z + c + 0.25 · p)
    return z, z_conv

TRAIN(images, identity labels, camera-view labels)
    for each minibatch with both views:
        a_A, a_G ← SVA(current prompts)
        z_i, z_conv_i ← FORWARD(x_i, s_i, a_A, a_G) for every image i
        L_native ← view-specific ID classification and triplet losses on z
        for each direction s→t in {A→G, G→A}:
            U ← converted descriptors z_conv from view s
            R ← stop_gradient(native descriptors z from view t)
            M ← mean of R for each identity present in view t
            H_fixed ← H_t with classifier and normalization fixed for TIC
            L_id ← ID classification of U with H_fixed and source labels
            L_tri ← directional triplet with U as anchors and R as references
            L_cos ← cosine loss from U to same-identity prototypes in M
            L_nce ← Proto-NCE from U to the prototypes in M
            L_TIC[s→t] ← weighted sum of L_id, L_tri, L_cos, and L_nce
        Backpropagate L_native + L_TIC[A→G] + L_TIC[G→A]
        Update the trainable visual parameters, prompts, DFD, and native heads

INFER(query images, gallery images, camera-view labels)
    a_A, a_G ← SVA(trained prompts)
    for every query and gallery image x with observed view s:
        z, z_conv ← FORWARD(x, s, a_A, a_G)
        if using flips, average each raw descriptor with its flipped counterpart
        slots[x] ← L2NormalizeEach(z, z_conv) if s = A
                   L2NormalizeEach(z_conv, z) if s = G
    for each query q and gallery image g:
        d_A, d_G ← cosine distances between corresponding slots of q and g
        d(q, g) ← d_A for A→A, d_G for G→G, (d_A + d_G) / 2 otherwise
    Rank each query's gallery by ascending distance

Q3: Ground-to-aerial setting mentioned in Sec. 4.1.

Thank you for pointing out the typo. Our CARGO evaluation follows ViSA’s [1] query–gallery protocol and reports the same four settings, ALL, A→A, G→G, and A→G. We mistakenly listed G→A as a separate CARGO setting in Section 4.1 and have corrected the manuscript.

Q4: Performance and generalization on AGReIDv2.

Thank you for this suggestion. To extend our evaluation beyond AGReIDv1 and CARGO, we have added experiments on AG-ReIDv2, an additional real-world benchmark covering aerial, CCTV, and wearable cameras. We evaluate all four retrieval directions between aerial images and the CCTV/wearable views, with the results presented in Table 2 below.

Meanwhile, TAFD is designed for cross-view identity matching within a dataset and does not include mechanisms specifically targeting cross-domain generalization. Following your suggestion, we evaluated the fixed CARGO-trained model directly on the same AG-ReIDv2 sets. The limited performance highlights a current limitation when transferring from synthetic images to a new real-world domain. However, this setting extends beyond the within-dataset cross-view matching problem studied in our work. We appreciate this suggestion and consider extending to address cross-domain generalization an important direction for future work.

Table 1. In-domain and cross-domain comparison on AG-ReIDv2 (map/Rank-1)

Training datasetA→CCTVCCTV→AA→WearableWearable→AAvg.
VISA83.61/89.4382.32/88.5785.99/91.6383.10/89.2387.76/89.72
AG-ReIDv283.98 / 88.9282.33 / 88.2487.20 / 92.0884.46 / 90.4384.49 / 89.92
CARGO9.25 / 14.434.47 / 6.4113.99 / 23.048.68 / 14.969.10 / 14.71

Q5. Does the real-world evidence support preservation of within-view discrimination?

Thanks for your thoughtful suggestion. Since AGReIDv1 is collected with one UAV camera and one CCTV camera. Within-view cross-camera retrieval requires each query to have a same-identity gallery image from a different camera. Under this requirement, the released images yield zero valid queries for both A→A and G→G. Among the benchmarks in our original evaluation, only synthetic CARGO supports both protocols. TAFD achieves 71.81% mAP and 77.50% Rank-1 for A→A, and 83.57% mAP and 90.18% Rank-1 for G→G.

However, from the method side, TAFD retains a native descriptor for each view, directly supervises it with identity classification and triplet losses, and uses it for within-view retrieval. During TIC, we detach the target-view native features, identity prototypes, and classifier weights. This lets the conversion losses train the transformed source features without directly updating the target references, which large maintain the feature under source-view.

Q6: Improvements consistency across repeated runs.

A: Thank you for insight comments. We agree that DFD provides the main performance gain, while TIC and SVA offer complementary refinements. TIC encourages identity preservation in the target view through additional training supervision, while SVA provides target-view guidance for deformation and feature aggregation. Our result is based on the average of multiple runs of the code. Specifically, we show standard deviation in the below table.

Table 1. Ablation study if TAFD.

SettingAvg. mAPAvg. Rank-1
+DFD80.08 ± 0.8386.78 ± 0.97
+DFD+TIC80.06 ± 0.6787.46 ± 0.78
+DFD+TIC+SV80.79 ± 0.6387.51 ± 1.00

[1] View-aware semantic alignment for aerial-ground person re-identification. CVPR 2026.

[2] View-decoupled transformer for person re-identification under aerial-ground camera network. CVPR 2024.

Finally, we hope our response has adequately addressed your concerns. If you have any further suggestions, please feel free to discuss them with us. We will provide the corresponding results as soon as possible.

I want to thank the authors for providing the detailed responses. However, concerns such as the assumption of true and complete camera label information and the marginal role of different components, as again pointed out in the new results, make the responses unconvincing; therefore, while I improved my rating, I feel it still requires sufficient work before full acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.