acceptodds
Under review as a conference paper at ICLR 2027

VR3D: View-Robust 3D Representation Learning for Aerial-Ground Person Re-Identification

Abstract

Aerial-ground person re-identification is a challenging task due to cross-platform viewpoint variations, which cause severe occlusion and geometric deformation. Existing methods attempt to learn view-invariant representations exclusively within the 2D image space, where drastic viewpoint variations cause the learned features to remain coupled with viewpoint bias. To address this, we propose VR3D, a View-Robust 3D Representation Learning framework that maps images into a unified 3D coordinate space to achieve view-independent feature interaction. Specifically, we introduce View-Robust 3D Representation Interaction, which leverages 3D priors extracted from single 2D observations to lift 2D appearance features into a canonical 3D space. VR3I employs 3D Geometry-Semantic Attention to establish interactions between 2D patches and 3D voxels from corresponding body parts based on their 3D spatial locations, effectively grounding 2D semantics within a 3D framework. In addition, as the reliability of these representations varies across samples due to viewpoint changes and 3D reconstruction errors, we introduce a Reliability-Aware Fusion that estimates sample-specific reliability and adaptively aggregates the multi-source representations. Extensive experiments on three benchmark datasets (CARGO, AG-ReID.v1, and AG-ReID.v2) demonstrate that VR3D outperforms recent methods, e.g., it achieves an improvement of 5.63% in Rank-1 on CARGO. Our code will be released.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Aerial-Ground Person Re-Identification

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2Uy5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the practical and challenging task of Aerial-Ground Person Re-Identification (AG-ReID), which involves severe scale, viewpoint, and geometric deformations across aerial and ground camera views. Existing AG-ReID methods primarily focus on 2D representation learning, limiting the mitigation of viewpoint bias in 2D space. To address this limitation, this paper proposes a View-Robust 3D Representation Learning (VR3D) framework that learns representations in a view-robust 3D space. Specifically, it maps image features into 3D human representations and employs View-Robust 3D Representation Interaction (VR3I) to establish interactions between 2D semantics and 3D human representations in a canonical 3D space, along with Reliability-Aware Fusion (RAF), which adaptively aggregates 2D, 3D, and interactive 2D–3D features. The proposed method is evaluated on three AG-ReID datasets to demonstrate its effectiveness.

Critical strengths
  1. The paper addresses a major limitation of existing AG-ReID methods, viewpoint bias in the 2D feature space learning, through View-Robust 3D Representation Learning.
  2. The paper's motivation to mitigate the bottleneck of 2D viewpoint bias is well-founded, and addressing this limitation by lifting representations into 3D space is interesting.
Critical weaknesses
  1. The paper relies on an offline 3D reconstruction pipeline, which limits the practical applicability of the proposed framework in real-world scenarios. The need to reconstruct 3D representations for each image introduces additional computational overhead, limiting the framework's suitability for end-to-end deployment.
  2. The overall contribution is limited, as the proposed VR3I and RAF modules are based on attention and fusion mechanisms that have been commonly explored in the Re-ID and computer vision literature.
  3. As the paper's main motivation or objective is to mitigate viewpoint bias, the use of view-specific features for Reliability-Aware Fusion requires stronger justification to demonstrate that these features reduce viewpoint-specific bias.
  4. The evaluation is limited, as the proposed method should be evaluated on large-scale real-world datasets such as LAGPeR and CP2108, along with cross-dataset evaluations to demonstrate its generalizability.
Questions for authors
  1. Provide a detailed computational complexity comparison, including key metrics such as FLOPs, trainable parameters, inference speed, and memory usage, against the ViT baseline and VDT method to analyze the practical applicability of the proposed method.
  2. Clarify the inference settings used during evaluation/testing, including how end-to-end retrieval is performed.
  3. As the paper's main motivation or objective is to mitigate viewpoint bias, the use of view-specific features for Reliability-Aware Fusion requires stronger justification to demonstrate that these features reduce viewpoint-specific bias.
8 comments

Weaknesses1

The paper relies on an offline 3D reconstruction pipeline, which limits the practical applicability of the proposed framework in real-world scenarios. The need to reconstruct 3D representations for each image introduces additional computational overhead, limiting the framework's suitability for end-to-end deployment.

R1

We sincerely thank the reviewer for this valuable comment. We would like to clarify that although VR3D employs SAM3/SAM3D for 3D feature extraction, this does not require repeated 3D reconstruction during retrieval. Similar to conventional embedding-based person ReID systems, VR3D performs feature extraction for gallery images only once during offline database construction. After encoding, only the final identity embeddings are retained, and all intermediate 3D features can be discarded. Therefore, online retrieval requires processing only the incoming query image, without reconstructing or re-encoding any gallery images.

Regarding end-to-end deployment, our measurements show that SAM3D reconstruction, the most GPU memory-intensive stage in our pipeline, has a peak memory usage of approximately 13.85 GiB, indicating that the pipeline can run on consumer-grade GPUs. Furthermore, the average intermediate 3D feature size is only 0.197 MB per image, and these features do not require long-term storage, avoiding additional storage overhead. For gallery construction, feature extraction and encoding can also be performed in parallel on cloud servers as a one-time offline process.

Overall, VR3D does not require repeated 3D reconstruction of gallery images during online retrieval. Instead, it incorporates the benefits of 3D structural information into a standard embedding-based retrieval framework. By combining offline gallery construction with flexible query processing, VR3D enables more viewpoint-robust identity representations while supporting practical deployment.

Weaknesses2

The overall contribution is limited, as the proposed VR3I and RAF modules are based on attention and fusion mechanisms that have been commonly explored in the Re-ID and computer vision literature.

R2

We sincerely thank the reviewer for this valuable comment. We would like to clarify that the core innovation of VR3D lies in introducing 3D representation learning into aerial-ground person ReID, providing a new approach to mitigating identity representation bias caused by extreme viewpoint variations. Attention and fusion serve as technical mechanisms to realize this design, rather than constituting our core contributions.

Specifically, existing methods mainly address viewpoint bias through 2D feature disentanglement. In contrast, our VR3I lifts 2D image patches from different viewpoints into a shared canonical 3D space, where geometric priors guide the attention computation to establish spatial associations between fine-grained 2D appearance features and 3D structural features. This enables the model to jointly exploit rich identity details from 2D images and structural information in 3D space, thereby learning more viewpoint-robust identity representations.

Building on this design, we further consider the influence of 3D reconstruction quality and viewing perspectives on feature reliability. Since the reliability of 2D appearance features, 3D structural features, and their interaction features varies across samples, we introduce the Reliability-Aware Fusion (RAF) module to adaptively adjust the fusion weights of these three representations according to sample-specific features. This allows the model to fully exploit their complementary information while reducing the influence of unreliable features on the final identity representation.

The experimental evidence further supports this distinction. The main ablation experiments in Table 4 validate the effectiveness of our core design, which consists of 3D representation learning and reliability-aware fusion for aerial-ground person ReID, rather than the use of attention and fusion operations alone.

Table 4. Ablation study of key components on AG-ReID.v1.“3D Global Feature” directly introduces the global 3D representation f3Df_{\text{3D}}, “Fused Feature” denotes the fuse representation ffusef_{\text{fuse}} learned by VR3I, and “Multi-Features” denotes average fusion of f2Df_{\text{2D}}, ffusef_{\text{fuse}}, and f3Df_{\text{3D}}. RAF further performs reliability-aware fusion.

VariantConfigurationA→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
A2D Baseline82.9174.4486.5978.57
B+ 3D Global Feature85.9676.6085.8678.62
C+ VR3I: Fused Feature86.0678.2087.5380.99
D+ VR3I: Multi-Features86.4479.3986.9082.18
E+ VR3I + RAF87.3979.9589.6082.87

Furthermore, the 3D-GSA ablation in Table 5 shows a substantial performance decrease when geometric priors are removed from the attention computation.

Table 5. Ablation study of interaction strategies in 3D-GSA on AG-ReID.v1. “Coarse” denotes fusion by adding global 2D and 3D features, while “Geometry” and “Semantic” use only geometric bias and semantic correlation, respectively. “Geo. + Sem.” combines both cues in 3D-GSA.

VariantInteractionA→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
ACoarse86.5378.9386.9081.15
BGeometry87.1179.6088.4682.15
CSemantic86.8279.4887.8481.74
DGeo. + Sem.87.3979.9589.6082.87

Moreover, with the rapid advances in 3D reconstruction, we believe that 3D representation learning will become increasingly valuable for addressing identity representation bias caused by extreme viewpoint variations. This research direction also enables aerial-ground person ReID to benefit directly from future advances in 3D reconstruction techniques.

Weaknesses3

As the paper's main motivation or objective is to mitigate viewpoint bias, the use of view-specific features for Reliability-Aware Fusion requires stronger justification to demonstrate that these features reduce viewpoint-specific bias.

R3

We sincerely thank the reviewer for this valuable question. We would like to clarify that introducing view-specific features is not intended to preserve viewpoint bias in the final identity representation. Instead, we leverage the additional visual cues provided by viewpoint information to assist in estimating the reliability of the 2D appearance features, 3D structural features, and their interaction features obtained through our 3D representation learning framework, thereby further enhancing the viewpoint robustness of the final identity representation.

Specifically, our method lifts images from different viewpoints into a shared canonical 3D space, where fine-grained 2D appearance information and 3D structural information are jointly exploited to mitigate viewpoint bias. However, the reliability of the three resulting feature representations varies across different samples. To address this issue, we introduce the Reliability-Aware Fusion (RAF) module. We further observe that different viewing perspectives affect 2D appearance features, 3D structural features, and their interaction features differently, causing their relative reliability to vary with viewpoint and 3D reconstruction quality. For example, under extreme overhead viewpoints, 2D images may lose important identity details, while reliable 3D structural features can still provide discriminative information. Therefore, although viewpoint information is a source of bias that needs to be mitigated in identity matching, it also provides useful visual cues that can assist in estimating the reliability of different feature representations. This observation motivates the introduction of view-specific information into RAF.

In our implementation, view-specific features are used only to construct the Query in the attention mechanism of RAF, providing supplementary information for estimating the relative reliability of the three feature branches. The Keys and Values are constructed from the 2D appearance features, 3D structural features, and their interaction features.

Thus, VR3I and RAF form a complementary and coherent design. VR3I mitigates viewpoint bias through 3D representation learning, while RAF further incorporates viewpoint information as an auxiliary cue to adaptively fuse identity features. Rather than simply removing all viewpoint information, we exploit its useful visual cues to support the learning of more viewpoint-robust identity representations.

Table 6. Ablation study of RAF on AG-ReID.v1. “Average” uses average fusion, “CA” performs cross-attention fusion, and “View” conditions fusion on view-related features.

VariantAverageCAViewA→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
A✓86.4479.3986.9082.18
B✓87.0179.5488.4682.15
C✓✓87.3979.9589.6082.87

The ablation results in Table 6 further support this design, showing that incorporating view-specific information helps the model better balance the contributions of the three feature representations, leading to improved retrieval performance. Moreover, the upper-right and lower-right examples in Fig. 7 illustrate that when extreme viewpoint variations degrade 2D appearance features while the reconstructed 3D structures remain reliable, RAF assigns higher fusion weights to the 3D structural branch.

Weaknesses4

The evaluation is limited, as the proposed method should be evaluated on large-scale real-world datasets such as LAGPeR and CP2108, along with cross-dataset evaluations to demonstrate its generalizability.

R4

We sincerely thank the reviewer for this valuable suggestion. To further evaluate the effectiveness and generalization ability of VR3D, we conduct additional experiments on the large-scale real-world LAGPeR and CP2108 datasets following their respective official cross-view retrieval evaluation protocols. We also conduct cross-dataset evaluations following the protocol of the original VDT paper.

(1) Evaluation on LAGPeR

LAGPeR is a large-scale real-world dataset, and we conduct additional experiments following its official evaluation protocols. As shown in the table, VR3D achieves the best performance across all three evaluation protocols. These results further demonstrate the effectiveness of VR3D on large-scale real-world aerial-ground person ReID datasets.

Comparison with existing methods on LAGPeR.

MethodA→G R1A→G mAPG→A R1G→A mAPG→(A+G) R1G→(A+G) mAPAverage R1Average mAP
ViT[a]38.6727.2532.0430.6918.8815.3129.8624.42
VDT[b]40.1528.9733.5531.9819.5016.4531.0725.80
SeCap[c]40.5129.5933.9532.6022.2317.9732.2326.72
SD-ReID[d]40.1529.7234.4732.8622.8518.8832.4927.15
ViSA[e]40.5830.0034.0132.7822.3218.3332.3027.04
VR3D41.6630.5835.6533.6923.2118.8933.5127.72

(2) Evaluation on CP2108

CP2108 is a large-scale real-world aerial-ground person ReID dataset, and we conduct additional experiments under cross-view retrieval protocols. As shown in the table, VR3D consistently achieves the best performance across all evaluation metrics. Compared with the strongest baseline, SVPR-ReID, VR3D improves Rank-1 accuracy by 5.28 and 4.67 percentage points under the A→G and G→A protocols, respectively, with corresponding mAP improvements of 2.18 and 2.25 percentage points. These results further demonstrate the effectiveness of VR3D in learning viewpoint-robust identity representations on large-scale real-world aerial-ground person ReID datasets.

Comparison with existing methods on CP2108.

MethodA→G R1A→G mAPG→A R1G→A mAPAverage R1Average mAP
VDT[b]42.7633.6440.5931.7441.6832.69
SeCap[c]63.5546.9563.3946.1963.4746.57
SVPR-ReID[f]64.2547.7263.7647.0464.0147.38
VR3D69.5349.9068.4349.2968.9849.60

(3) Cross-Dataset Generalization

To further evaluate the cross-dataset generalization ability of VR3D, we conduct additional experiments by transferring from CARGO to AG-ReID, following the evaluation protocol reported in Section 7.3, Table 4 of the original VDT paper. As shown in the table, VR3D achieves 29.89% Rank-1 and 18.04% mAP under the A→G protocol, and 20.89% Rank-1 and 15.63% mAP under the G→A protocol. Compared with VDT, VR3D improves these four metrics by 10.56, 6.23, 5.51, and 3.90 percentage points, respectively. The results demonstrate that our proposed viewpoint-robust 3D representation learning also substantially improves cross-dataset generalization ability.

MethodA→G Rank-1 (%)A→G mAP (%)G→A Rank-1 (%)G→A mAP (%)
ViT[a]1.591.953.012.31
VDT[b]19.3311.8115.3811.73
VR3D29.8918.0420.8915.63
Reviewer_A9YHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VR3D, which lifts 2D appearance features into a canonical 3D space for view-robust aerial-ground person ReID, using geometry-semantic attention and reliability-aware fusion to improve cross-view matching.

Critical strengths

The paper has a clear narrative, well-defined motivation, and strong experimental results.

Critical weaknesses
  1. The proposed method relies heavily on SAM3 and SAM3D and assumes reliable 3D reconstruction from a single 2D image, yet the associated risks, computational costs, runtime, and storage requirements are not sufficiently quantified.
  2. Although modeling interactions in 3D directly addresses the extreme viewpoint variations in cross-view aerial-ground ReID, the methodological novelty is limited, and comparisons with the most relevant 3D-based competing methods are missing.
Questions for authors

My main concern is whether the clear experimental gains justify the additional computational and storage costs associated with SAM3/SAM3D compared with existing aerial-ground ReID methods.

4 comments

Weaknesses1

The proposed method relies heavily on SAM3 and SAM3D and assumes reliable 3D reconstruction from a single 2D image, yet the associated risks, computational costs, runtime, and storage requirements are not sufficiently quantified.

R1

We sincerely thank the reviewer for this valuable suggestion.

About the Reliance on SAM3/SAM3D and the Risks of Unreliable 3D Reconstruction

First, we clarify that VR3D does not assume that SAM3/SAM3D produces fully reliable 3D reconstructions for every input image. Instead, recognizing that single-image 3D reconstruction inevitably carries the risk of reconstruction errors, we introduce a Reliability-Aware Fusion (RAF) module after learning three complementary representations in a shared canonical 3D space. This module adaptively adjusts the contribution of each feature branch to the final identity representation.

Specifically, RAF does not unconditionally rely on reconstructed 3D geometric information. Instead, it dynamically assigns fusion weights. When the reconstructed 3D structure is unreliable, the model can reduce its reliance on distorted 3D structural features and assign greater weight to the original 2D appearance features, enabling adaptive feature weighting at the sample level.

Fig. 10(a) provides a clear illustration of this mechanism. As shown in the upper-left example, when severe 3D reconstruction distortion occurs, the model assigns a lower fusion weight to the 3D structural branch while substantially increasing the weight of the original 2D appearance branch. This example demonstrates that VR3D can adaptively adjust the contributions of different representations to reduce the influence of distorted 3D information on the final identity representation.

Therefore, the core design of VR3D does not depend on obtaining accurate 3D reconstructions for all samples. Instead, reliability-aware fusion enables adaptive weighting across different feature branches. When 3D structural information becomes unreliable, the model can place greater emphasis on relatively reliable 2D appearance features, thereby mitigating the impact of reconstruction errors.

About Computational Efficiency

We first clarify that VR3D follows a standard embedding-based retrieval framework, similar to existing ReID methods. SAM3/SAM3D feature extraction for gallery images is performed only once during offline gallery construction, and the resulting identity embeddings are cached for subsequent retrieval. Therefore, online retrieval only requires processing the incoming query and computing its similarity with the precomputed gallery embeddings, without reconstructing or re-encoding the gallery. We provide detailed measurements of preprocessing efficiency, storage requirements and VR3D network complexity below.

(1) Deployment and Offline Preprocessing

During gallery construction, SAM3 and SAM3D extract intermediate 3D features from each image. These features, together with the original 2D image, are fed into VR3D to generate the final identity embedding. Once encoding is complete, the intermediate 3D features can be discarded, and only the final identity embedding is retained. Thus, like conventional image-based ReID methods, VR3D maintains a gallery database consisting solely of identity embeddings.

On a single NVIDIA A100 40 GB GPU, our 3D feature extraction pipeline processes approximately 637 images per hour. In the future, inference acceleration techniques such as Fast-SAM3D could further improve the efficiency of offline gallery construction, potentially increasing the processing throughput to approximately 1,690 images per hour.Importantly, this preprocessing is performed only once during offline gallery construction and does not contribute to the online retrieval latency of new queries.

(2) Storage Requirements.

VR3D directly utilizes intermediate 3D structural features generated during SAM3D reconstruction, without requiring the complete reconstructed 3D models to be stored. To quantify the intermediate feature storage requirements, we randomly sample 10,000 images from each of the three datasets, totaling 30,000 samples. The average intermediate feature size is approximately 0.197 MB per image. These features are only needed during identity embedding generation and can be discarded after encoding. Therefore, VR3D does not introduce additional long-term storage requirements for 3D features, and the final gallery database contains only identity embeddings.

(3) Computational Efficiency of VR3D

We report the inference-time computational complexity of VR3D on AG-ReID.v1 in the following table.

MethodParams (M)GFLOPs / ImageFPS (B64)Peak Memory (MiB)A→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
ViT[a]85.7522.681932.4564181.2872.3882.6473.35
VDT[b]85.7522.861917.1164582.9174.4486.5978.57
SeCap[c]130.9747.70944.15116484.0376.1687.0178.34
ViSA[d]271.7742.06453.03244084.8175.9586.3878.24
VR3D(Ours)107.4142.93481.8498187.3979.9589.6082.87

As shown in the table, the substantial improvements in retrieval performance demonstrate that VR3D achieves a favorable trade-off between accuracy and computational efficiency.

Overall, VR3D combines more discriminative and viewpoint-robust identity representations with a practical embedding-based retrieval framework. The gallery requires only one-time offline feature extraction, and the final database stores only identity embeddings without additional long-term 3D feature storage. This design allows VR3D to benefit from 3D representation learning while preserving the efficiency of standard single-vector retrieval.

Weaknesses2

Although modeling interactions in 3D directly addresses the extreme viewpoint variations in cross-view aerial-ground ReID, the methodological novelty is limited, and comparisons with the most relevant 3D-based competing methods are missing.

R2

We sincerely thank the reviewer for this valuable comment. We would like to emphasize that the core innovation of VR3D lies in introducing 3D representation learning to address identity representation bias caused by extreme viewpoint variations in aerial-ground person ReID. By jointly leveraging fine-grained 2D appearance details and 3D structural information in a shared canonical 3D space, VR3D learns more complete and viewpoint-robust identity representations.

Specifically, existing aerial-ground person ReID methods mainly address viewpoint bias through view disentanglement (e.g., VDT[b]), learning disentangled features (e.g., SeCap[c] and ViSA[d]), or incorporating additional attribute-based semantic information. However, extreme viewpoint variations between aerial and ground cameras lead to substantial changes in visible body regions, appearance details, and spatial relationships. Relying solely on 2D feature learning makes it difficult to fully exploit the geometric consistency of human body structures across different viewpoints. We observe that 3D spatial representations are inherently more robust to viewpoint changes. Based on this insight, we aim to learn more viewpoint-robust 3D representations in a canonical 3D space.

Our method is clearly distinguished from existing 3D-based ReID approaches. OG-Net[e] primarily focuses on ground-based person ReID scenarios, leveraging 3D point clouds to learn human structural representations rather than specifically addressing viewpoint bias by mapping images from different viewpoints into a shared canonical 3D space. In contrast, 3D-LENS[f] mainly employs 3D reconstruction to synthesize images from different viewpoints for training data augmentation. Unlike these methods, VR3D further lifts 2D image patches from different viewpoints into a shared canonical 3D space, where it jointly exploits rich fine-grained identity details from 2D images and the geometric consistency of 3D structural information across viewpoints. This enables VR3D to learn more complete and viewpoint-robust 3D identity representations, directly mitigating appearance bias caused by extreme aerial-ground viewpoint variations.

Furthermore, considering that different viewing perspectives and 3D reconstruction quality affect the reliability of 2D appearance features, 3D structural features, and their interaction features, we introduce the Reliability-Aware Fusion (RAF) module. RAF adaptively adjusts the fusion weights of these three feature branches according to their relative reliability for each sample, effectively exploiting the complementary strengths of different representations.

Moreover, with the rapid advances in 3D reconstruction, we believe that 3D representation learning will become increasingly valuable for addressing representation bias caused by extreme viewpoint variations. This research direction also enables aerial-ground person ReID to benefit from future advances in 3D reconstruction techniques. We hope these responses resolve the initial concerns. We would be grateful if you could consider raising the score to support our work. We would be glad to clarify any remaining points if you have any further questions.

Question1

My main concern is whether the clear experimental gains justify the additional computational and storage costs associated with SAM3/SAM3D compared with existing aerial-ground ReID methods.

R3

We sincerely thank the reviewer for this valuable suggestion. We first clarify that VR3D follows a standard embedding-based retrieval framework, similar to existing ReID methods. SAM3/SAM3D feature extraction for gallery images is performed only once during offline gallery construction, and the resulting identity embeddings are cached for subsequent retrieval. The intermediate 3D features can be discarded after generating the final identity embeddings and do not require long-term storage. Therefore, online retrieval only requires processing the incoming query and computing its similarity with the precomputed gallery embeddings, without reconstructing or re-encoding the gallery. We provide detailed measurements of preprocessing efficiency, storage requirements, and VR3D network complexity below.

(1) Deployment and Offline Preprocessing

During gallery construction, SAM3 and SAM3D extract intermediate 3D features from each image. These features, together with the original 2D image, are fed into VR3D to generate the final identity embedding. Once encoding is complete, the intermediate 3D features can be discarded, and only the final identity embedding is retained. Thus, like conventional image-based ReID methods, VR3D maintains a gallery database consisting solely of identity embeddings.

On a single NVIDIA A100 40 GB GPU, our 3D feature extraction pipeline processes approximately 637 images per hour. Importantly, this preprocessing is performed only once during offline gallery construction, and the gallery does not need to be reconstructed or updated when processing new queries.

(2) Storage Requirements

VR3D directly utilizes intermediate 3D structural features generated during SAM3D reconstruction, without requiring the complete reconstructed 3D models to be stored. To quantify the intermediate feature storage requirements, we randomly sample 10,000 images from each of the three datasets, totaling 30,000 samples. The average intermediate feature size is approximately 0.197 MB per image. These features are only needed during identity embedding generation and can be discarded after encoding. Therefore, VR3D does not introduce additional long-term storage requirements for 3D features, and the final gallery database contains only identity embeddings.

(3) Computational Efficiency of VR3D

We report the inference-time computational complexity of VR3D on AG-ReID.v1 in the following table.

MethodParams (M)GFLOPs / ImageFPS (B64)Peak Memory (MiB)A→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
ViT[a]85.7522.681932.4564181.2872.3882.6473.35
VDT[b]85.7522.861917.1164582.9174.4486.5978.57
SeCap[c]130.9747.70944.15116484.0376.1687.0178.34
ViSA[d]271.7742.06453.03244084.8175.9586.3878.24
VR3D(Ours)107.4142.93481.8498187.3979.9589.6082.87

As shown in the table, the substantial improvements in retrieval performance demonstrate that VR3D achieves a favorable trade-off between accuracy and computational efficiency.

Overall, VR3D combines more discriminative and viewpoint-robust identity representations with a practical embedding-based retrieval framework. The gallery requires only one-time offline feature extraction, and the final database stores only identity embeddings without additional long-term 3D feature storage. This design allows VR3D to benefit from 3D representation learning while preserving the efficiency of standard single-vector retrieval. Therefore, we believe that the substantial improvements in retrieval performance justify the additional computational costs associated with SAM3/SAM3D.

We hope these responses resolve the initial concerns. We would be grateful if you could consider raising the score to support our work. We would be glad to clarify any remaining points if you have any further questions.

[a] Dosovitskiy A. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020.

[b] Zhang Q, Wang L, Patel V M, et al. View-decoupled transformer for person re-identification under aerial-ground camera network[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 22000-22009.

[c] Wang S, Wang Y, Wu R, et al. SeCap: Self-calibrating and adaptive prompts for cross-view person re-identification in aerial-ground networks[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 22119-22128.

[d] Zhang Q, Cai Z, Zhao P, et al. View-Aware Semantic Alignment for Aerial-Ground Person Re-Identification[J]. arXiv preprint arXiv:2605.18192, 2026.

[e] Zheng Z, Wang X, Zheng N, et al. Parameter-efficient person re-identification in the 3D space[J]. IEEE Transactions on Neural Networks and Learning Systems, 2022, 35(6): 7534-7547.

[f] Grolleau W, Sabourin A, Lapouge G, et al. 3D-LENS: A 3D Lifting-Based Elevated Novel-View Synthesis Method for Single-View Aerial-Ground Re-Identification[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026: 205-221.

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

In response to the significant variations in viewpoint encountered in pedestrian re-identification in open-space scenarios, this paper proposes a framework that integrates two-dimensional appearance, three-dimensional geometric interaction and viewpoint-conditional fusion. The research motivation is clear and the overall design is coherent. The main retrieval results on the three benchmarks are competitive, and component ablation studies provide support for the practical value of geometric and semantic interaction, as well as dynamic fusion. Consequently, the paper possesses considerable methodological value and application potential.

Critical strengths
  1. Geometric distance and semantic relevance jointly guide 2D–3D interactions, taking into account both identity appearance details and structural information, and are well-suited to cross-view ReID tasks
  2. CARGO A↔G achieves a Rank-1/mAP of 59.38/54.57, outperforming ViSA in the table by 5.63/4.93 percentage points; multi-protocol gains are also observed on other benchmarks.
  3. Tables 4–6 respectively evaluate 3D information, geometric and semantic interaction, and viewpoint-conditional fusion, supporting the practical effectiveness of the components.
Critical weaknesses
  1. The rigidity and scale transformation in Equation 6 do not, in themselves, guarantee the correspondence of body parts across different orientations, limb postures and reconstruction errors; experiments involving correspondence errors or coordinate perturbations are required.
  2. Table 6 demonstrates that fusion is useful, but the selected visualisations do not yet prove the quality of weight-tracking reconstruction or branching errors.
  3. There is a lack of variance across multiple random seeds, as well as data on pre-processing time, memory usage, failure rates and new query latency for SAM3/SAM3D.
Questions for authors

Details in Strengths and Weaknesses.

4 comments

Weaknesses1

The rigidity and scale transformation in Equation 6 do not, in themselves, guarantee the correspondence of body parts across different orientations, limb postures and reconstruction errors; experiments involving correspondence errors or coordinate perturbations are required.

R1

About Equation 6

We sincerely thank the reviewer for this valuable suggestion. This concern directly relates to a key design of our method: transforming fine-grained 2D image features from different viewpoints into a canonical 3D space, where they can be geometrically aligned with 3D structural features to learn viewpoint-robust 3D representations under geometric constraints.

Specifically, Eq. (6) defines the transformation from the original 3D coordinate system to the canonical 3D coordinate system:

P^m=(Pm−T)RTs\hat{P}_m = \frac{(P_m - T)R^{\mathrm{T}}}{s}

Here, Pₘ denotes the 3D coordinates of an image patch predicted during SAM3D inference, while R, T, and s represent the rotation matrix, translation vector, and scale factor, respectively. Through translation compensation, inverse rotation, and scale normalization, this transformation compensates for global pose and scale differences introduced by different viewing perspectives. It maps the 2D image patches into a canonical 3D coordinate system that shares a consistent orientation and spatial reference with the extracted 3D structural features. Consequently, fine-grained image features and 3D structural features can establish geometric associations within the same coordinate system.

Building on this geometric reference, 3D-GSA incorporates the spatial distances between canonical 3D coordinates as geometric priors into the attention computation, together with the semantic similarities between 2D appearance features and 3D structural features. Importantly, even when the estimated 3D coordinates contain geometric errors, semantic similarity provides complementary information for feature association. By jointly considering geometric proximity and semantic correspondence, 3D-GSA enables more effective interactions between fine-grained appearance details and the corresponding 3D structural information of human body regions.

The visualizations in Figs. 6 and 9 of Appendix C further illustrate how 2D image patches are lifted into the same canonical 3D space as the extracted 3D structural features, enabling spatially guided associations between fine-grained image semantics and 3D structural information. Furthermore, the ablation results in Table 5 show that removing the geometric priors leads to a performance drop, demonstrating the importance of accurate geometric associations for learning viewpoint-robust 3D representations.

Table 5. Ablation study of interaction strategies in 3D-GSA on AG-ReID.v1. “Coarse” denotes fusion by adding global 2D and 3D features, while “Geometry” and “Semantic” use only geometric bias and semantic correlation, respectively. “Geo. + Sem.” combines both cues in 3D-GSA.

VariantInteractionA→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
ACoarse86.5378.9386.9081.15
BGeometry87.1179.6088.4682.15
CSemantic86.8279.4887.8481.74
DGeo. + Sem.87.3979.9589.6082.87

About Random Coordinate Perturbation Experiments

To further validate the importance of accurate geometric associations in constructing effective geometric priors, we conduct additional random coordinate perturbation experiments by perturbing the 3D coordinates of 50% and 100% of the samples, respectively. The results show that retrieval performance decreases as the geometric associations are disrupted. This further demonstrates the effectiveness of our geometric coordinate transformation in establishing meaningful geometric associations, which facilitate more effective fine-grained interactions between 2D appearance features and 3D structural features.

Coordinate PerturbationA→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
0% (Original)87.3979.9589.6082.87
50% Random Coordinate86.7279.4387.3281.64
100% Random Coordinate86.2578.4787.1181.18

Weaknesses2

Table 6 demonstrates that fusion is useful, but the selected visualizations do not yet prove the quality of weight-tracking reconstruction or branching errors.

R2

We sincerely thank the reviewer for this valuable suggestion. We first clarify that the key role of RAF is to adaptively adjust the fusion weights of 2D appearance features, 3D structural features, and their interaction features according to their relative reliability for each sample.

As shown in Fig. 7, different samples exhibit distinct fusion weight distributions. In the upper-left example, both the 2D appearance and the reconstructed 3D structure are relatively unclear, resulting in similar weights across the three branches. In the lower-left example, both the 2D appearance details and the reconstructed 3D structure are clearly preserved. Consequently, the interaction features receive a higher fusion weight, indicating that RAF effectively exploits complementary information from both modalities to refine the identity representation. In the upper-right and lower-right examples, severe viewpoint variations reduce the reliability of 2D appearance features, while the reconstructed 3D structures remain relatively complete and preserve recognizable identity characteristics. Accordingly, RAF assigns higher weights to the 3D structural branch.

Furthermore, in the extreme case shown in the upper-left example of Fig. 10(a), SAM3D incorrectly reconstructs a person as an animal, resulting in severe distortion of the 3D structural information. In this case, RAF substantially reduces the fusion weight of the 3D branch, preventing unreliable 3D information from dominating the final identity representation.

These representative examples demonstrate how RAF dynamically adjusts fusion weights according to the reliability of different feature branches. They also highlight the importance of reliability-aware fusion in maintaining discriminative identity representations when 3D reconstruction errors occur.

Weaknesses3

There is a lack of variance across multiple random seeds, as well as data on pre-processing time, memory usage, failure rates and new query latency for SAM3/SAM3D.

R3

We sincerely thank the reviewer for this valuable suggestion.

About Random Seed Experiments

To evaluate the stability of VR3D, we conduct three independent runs on AG-ReIDv1 and AG-ReIDv2 using random seeds 0, 42, and 123. We report the results for each seed, together with the mean and standard deviation across the three runs under all evaluation protocols. The results exhibit only minor variations across different random seeds, as reflected by the relatively small standard deviations, demonstrating the stability of VR3D across independent training runs.

Agreidv1

MethodA→G Rank-1 (%)A→G mAP (%)G→A Rank-1 (%)G→A mAP (%)
VDT[a]82.9174.4486.5978.57
SeCap[b]84.0376.1687.0178.34
SD-ReID[c]85.1675.4085.9777.02
SVPR-ReID[d]85.3477.8587.3280.55
VR3D (Ours)87.30 ± 0.1679.85 ± 0.2189.15 ± 0.3982.75 ± 0.18
Random SeedA→G Rank-1A→G mAPG→A Rank-1G→A mAP
Seed 087.3979.9589.6082.87
Seed 4287.1179.9988.9882.54
Seed 12387.3979.6088.8882.84
Mean ± Std87.30 ± 0.1679.85 ± 0.2189.15 ± 0.3982.75 ± 0.18

Agreidv2

MethodA→C R1 (%)A→C mAP (%)A→W R1 (%)A→W mAP (%)C→A R1 (%)C→A mAP (%)W→A R1 (%)W→A mAP (%)
SeCap[b]88.1280.8491.4484.0188.2479.9987.5680.15
SD-ReID[c]87.0480.6190.8684.0686.7479.2486.7980.12
ViSA[e]89.4383.6191.6385.9988.5782.3289.2383.10
VR3D(Ours)90.15 ± 0.2084.49 ± 0.0793.68 ± 0.1487.96 ± 0.1090.69 ± 0.1484.28 ± 0.1489.76 ± 0.3984.07 ± 0.11
Random SeedA→C R1 (%)A→C mAP (%)A→W R1 (%)A→W mAP (%)C→A R1 (%)C→A mAP (%)W→A R1 (%)W→A mAP (%)
Seed 089.9884.5493.5787.9690.5684.3989.7084.17
Seed 4290.1184.5193.6287.8690.6784.1389.4083.95
Seed 12390.3784.4193.8488.0590.8384.3390.1784.09
Mean ± Std90.15 ± 0.2084.49 ± 0.0793.68 ± 0.1487.96 ± 0.1090.69 ± 0.1484.28 ± 0.1489.76 ± 0.3984.07 ± 0.11

About Computational Efficiency

(1) Deployment and Retrieval Workflow

We first clarify the practical deployment and retrieval workflow of VR3D. For gallery construction, SAM3 and SAM3D are used to extract intermediate 3D features offline. These features, together with the original images, are then fed into VR3D to generate the final identity embeddings. This process is performed only once during gallery construction. After encoding, the intermediate 3D features can be discarded, and only the final identity embeddings need to be retained. Therefore, similar to conventional image-based ReID methods, VR3D ultimately maintains a gallery database consisting only of identity embeddings.

(2) Data Preprocessing

On a single NVIDIA A100 40 GB GPU, our current 3D feature extraction pipeline processes approximately 637 images per hour. In the future, inference acceleration techniques such as Fast-SAM3D could further improve the efficiency of offline gallery construction, potentially increasing the processing throughput to approximately 1,690 images per hour. Importantly, gallery feature extraction is performed only once offline and does not contribute to the online retrieval latency of a new query. This makes the proposed workflow practical for retrieval applications.

(3) Storage Cost

VR3D does not require storing the complete 3D Gaussian models generated by SAM3D. Instead, it directly utilizes intermediate 3D features produced during the reconstruction process. Specifically, the required data consist only of the coordinates of image patches in the canonical 3D space, the voxel coordinates of 3D structural features, and their corresponding intermediate 3D features. Since the number of voxels varies across samples, the temporary feature size also varies. We randomly sample 10,000 images from each of the three datasets, resulting in 30,000 samples in total. The average intermediate feature size is approximately 0.197 MB per image. More importantly, these features are used only to generate the final identity embeddings and can be discarded immediately after encoding. Therefore, VR3D introduces no additional long-term storage requirements for 3D features.

(4) New Query Latency

VR3D follows a standard embedding-based retrieval framework, where gallery images are processed only once during offline database construction, and only the final identity embeddings are retained. Therefore, online retrieval does not require reprocessing any gallery images. For a previously unseen query, the system only needs to perform SAM3/SAM3D feature extraction and VR3D encoding once, followed by similarity computation against the precomputed gallery embeddings. Our measurements show that the average end-to-end retrieval latency for a new query is approximately 9.2 s. For queries whose identity embeddings are already cached, retrieval requires only vector similarity computation at the millisecond level, without additional neural network inference. Importantly, the processing latency for previously unseen images is incurred only once, as their identity embeddings are subsequently cached and directly reused for similarity computation. Therefore, we believe that this one-time processing cost is reasonable for practical retrieval applications.

Leveraging large pretrained models to enhance retrieval representations in ReID has also been explored in recent studies. For example, IDEA[f] (CVPR 2025) and CoT-ReID[g] (CVPR 2026) employ Qwen-VL to generate semantic descriptions and reasoning information. We further reproduce IDEA-ReID and CoT-ReID using Qwen3.8-27B to compare their inference time and peak GPU memory usage. As shown in the table, VR3D achieves lower inference latency and peak GPU memory usage than the two methods.

MethodInference Time (s)Peak GPU Memory
IDEA-ReID[f]11.5059.03 GiB
CoT-ReID[g]156.2459.03 GiB
VR3D(Ours)9.2013.85 GiB

(5) Computational Efficiency of VR3D

We report the inference-time computational complexity of VR3D on AG-ReID.v1 in the following table.

MethodParams (M)GFLOPs / ImageFPS (B64)Peak Memory (MiB)A→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
ViT[h]85.7522.681932.4564181.2872.3882.6473.35
VDT[a]85.7522.861917.1164582.9174.4486.5978.57
SeCap[b]130.9747.70944.15116484.0376.1687.0178.34
ViSA[e]271.7742.06453.03244084.8175.9586.3878.24
VR3D(Ours)107.4142.93481.8498187.3979.9589.6082.87

As shown in the table, the substantial improvements in retrieval performance demonstrate that VR3D achieves a favorable trade-off between accuracy and computational efficiency.

(6)Reconstruction Failures and Reliability

Our 3D feature extraction pipeline consists of two main stages: object segmentation using SAM3 and 3D reconstruction using SAM3D.

SAM3 Segmentation We use the person prompt to generate candidate pedestrian masks and select the most appropriate mask based on the target's position and area. We additionally introduce the bag prompt to include backpacks and other carried items relevant to identity recognition in the target region, thereby obtaining a more complete pedestrian mask. This strategy substantially improves the quality of target segmentation, and segmentation failures are rarely observed in practice.

SAM3D Reconstruction We feed the original image and the selected target mask into SAM3D to reconstruct the 3D structure and extract intermediate 3D features. Although the reconstruction process produces 3D features for each input, manual inspection reveals that low-resolution images can lead to distorted 3D reconstructions. This limitation is precisely one of the key motivations for introducing the Reliability-Aware Fusion (RAF) module alongside 3D representation learning. RAF adaptively adjusts the fusion weights according to the relative reliability of different feature branches, reducing the influence of unreliable 3D features while effectively exploiting the complementary information between 2D appearance and 3D structure.

We hope these responses resolve the initial concerns. We would be grateful if you could consider raising the score to support our work. We would be glad to clarify any remaining points if you have any further questions.

[a] Zhang Q, Wang L, Patel V M, et al. View-decoupled transformer for person re-identification under aerial-ground camera network[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 22000-22009.

[b] Wang S, Wang Y, Wu R, et al. SeCap: Self-calibrating and adaptive prompts for cross-view person re-identification in aerial-ground networks[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 22119-22128.

[c] Hu X, Zhang P, Wang Y, et al. View-aware stable diffusion for aerial-ground person re-identification[J]. arXiv preprint arXiv:2504.09549, 2025, 3.

[d] Zheng A, Xie H, Wan X, et al. Semantic-Driven Visual Progressive Refinement for Aerial-Ground Person ReID: A Challenging Large-Scale Benchmark[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2026, 40(16): 13360-13368.

[e] Zhang Q, Cai Z, Zhao P, et al. View-Aware Semantic Alignment for Aerial-Ground Person Re-Identification[J]. arXiv preprint arXiv:2605.18192, 2026.

[f]Wang Y, Lv Y, Zhang P, et al. Idea: Inverted text with cooperative deformable aggregation for multi-modal object re-identification[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 29701-29710.

[g]Gao Y, Li S, Liu Z, et al. Chain-of-Thought Guided Multi-Modal Object Re-Identification[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 37705-37714.

[h] Dosovitskiy A. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020.

Comments on the paper

2 comments

(3) Clarification of Our Contributions (A9YH, 2Uy5)

We would like to emphasize that the core contribution of VR3D lies in introducing 3D representation learning into Aerial-Ground ReID to address identity representation bias caused by extreme viewpoint variations.

Specifically, existing methods mainly mitigate viewpoint bias through 2D feature disentanglement. In contrast, our method maps 2D image patches from different viewpoints into a shared canonical 3D space, where fine-grained interactions between 2D appearance features and 3D structural features are jointly guided by geometric and semantic information. This design preserves rich identity details from the original 2D images while incorporating 3D structural information. Furthermore, we observe that the reliability of 2D appearance features, 3D structural features, and their interaction features varies across samples due to differences in viewpoints and reconstruction quality. Therefore, we introduce the Reliability-Aware Fusion (RAF) module to adaptively adjust the contributions of these three feature branches. Attention serves as a technical mechanism for implementing our proposed design rather than constituting the core contribution itself. The ablation experiments further validate the effectiveness of these designs.

[a] Dosovitskiy A. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020.

[b] Zhang Q, Wang L, Patel V M, et al. View-decoupled transformer for person re-identification under aerial-ground camera network[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 22000-22009.

[c] Wang S, Wang Y, Wu R, et al. SeCap: Self-calibrating and adaptive prompts for cross-view person re-identification in aerial-ground networks[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 22119-22128.

[d] Zhang Q, Cai Z, Zhao P, et al. View-Aware Semantic Alignment for Aerial-Ground Person Re-Identification[J]. arXiv preprint arXiv:2605.18192, 2026.

[e] Hu X, Zhang P, Wang Y, et al. View-aware stable diffusion for aerial-ground person re-identification[J]. arXiv preprint arXiv:2504.09549, 2025, 3.

[f] Zheng A, Xie H, Wan X, et al. Semantic-Driven Visual Progressive Refinement for Aerial-Ground Person ReID: A Challenging Large-Scale Benchmark[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2026, 40(16): 13360-13368.

We sincerely thank all reviewers for their constructive comments and recognition of our research contributions. These valuable suggestions help us further clarify the research motivation, technical contributions, and practical applicability of VR3D.

Reviewers sGho, A9YH, and 2Uy5 all recognize the clear motivation of introducing viewpoint-robust 3D representation learning to address 2D viewpoint bias in aerial-ground person ReID. In particular, reviewer sGho notes that geometric distance and semantic relevance jointly guide fine-grained 2D–3D interactions in a canonical 3D space, effectively combining identity appearance details with structural information for cross-view ReID. Reviewer 2Uy5 further considers the idea of mapping 2D features into 3D space to mitigate viewpoint bias interesting. In addition, reviewers sGho and A9YH recognize the strong retrieval performance of VR3D across multiple benchmarks.

To further address the reviewers' concerns, we provide additional evidence regarding the practical retrieval workflow and computational efficiency, competitive performance on the large-scale LAGPeR and CP2108 datasets, and substantial improvements in cross-dataset generalization. We also further clarify our core technical contributions.

Common Concerns

(1) Retrieval Pipeline and Computational Efficiency (sGho, A9YH, 2Uy5)

We clarify that VR3D follows a standard embedding-based retrieval framework. During gallery construction, SAM3/SAM3D feature extraction and VR3D encoding are performed only once offline. After encoding, the intermediate 3D features are discarded, and only the final identity embeddings are retained. Therefore, online retrieval does not require reconstructing or re-encoding gallery images.

We report the computational complexity of VR3D during inference on AG-ReID.v1. VR3D achieves substantial performance improvements under both evaluation protocols while maintaining competitive inference throughput.

MethodParams (M)GFLOPs / ImageFPS (B64)Peak Memory (MiB)A→G R1 (%)A→G mAP (%)G→A R1 (%)G→A mAP (%)
ViT[a]85.7522.681932.4564181.2872.3882.6473.35
VDT[b]85.7522.861917.1164582.9174.4486.5978.57
SeCap[c]130.9747.70944.15116484.0376.1687.0178.34
ViSA[d]271.7742.06453.03244084.8175.9586.3878.24
VR3D(Ours)107.4142.93481.8498187.3979.9589.6082.87

(2) Large-Scale Evaluation and Cross-Dataset Generalization (2Uy5)

About Large-Scale Evaluation

LAGPeR and CP2108 are large-scale real-world aerial-ground person ReID datasets. We conduct additional experiments following their respective official evaluation protocols. Among the compared methods, VR3D achieves the best performance across all evaluated cross-view retrieval protocols on both datasets.

LAGPeR

MethodA→G R1A→G mAPG→A R1G→A mAPG→(A+G) R1G→(A+G) mAPAverage R1Average mAP
ViT[a]38.6727.2532.0430.6918.8815.3129.8624.42
VDT[b]40.1528.9733.5531.9819.5016.4531.0725.80
SeCap[c]40.5129.5933.9532.6022.2317.9732.2326.72
SD-ReID[e]40.1529.7234.4732.8622.8518.8832.4927.15
ViSA[d]40.5830.0034.0132.7822.3218.3332.3027.04
VR3D (Ours)41.6630.5835.6533.6923.2118.8933.5127.72

CP2108

MethodA→G R1A→G mAPG→A R1G→A mAPAverage R1Average mAP
VDT[b]42.7633.6440.5931.7441.6832.69
SeCap[c]63.5546.9563.3946.1963.4746.57
SVPR-ReID[f]64.2547.7263.7647.0464.0147.38
VR3D (Ours)69.5349.9068.4349.2968.9849.60

Cross-Dataset Generalization

We further conduct cross-dataset evaluations by transferring from CARGO to AG-ReID, following the evaluation protocol of the original VDT paper. The results demonstrate that our proposed viewpoint-robust 3D representation learning substantially improves cross-dataset generalization ability for aerial-ground person ReID.

MethodA→G Rank-1 (%)A→G mAP (%)G→A Rank-1 (%)G→A mAP (%)
ViT[a]1.591.953.012.31
VDT[b]19.3311.8115.3811.73
VR3D(Ours)29.8918.0420.8915.63

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.