acceptodds
Withdrawn from ICLR 2027

Predict Before You Match: Cross-View Predictive Learning for Aerial-Ground Person Re-Identification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The Aerial-to-Ground Person Re-identification (AG-ReID) task requires identity matching across observation data with significant visual feature discrepancies. Existing methods primarily address these differences through view-invariant or view-aware representation learning. This paper explores a complementary approach: although the visual appearance of the same identity varies across views, predictable correlations persist between them. This implies that, beyond direct feature alignment, cross-view predictive relationships can serve as a latent supervisory signal for identity. Based on this insight, we propose an identity-conditioned cross-view latent prediction method, Predictive Re-identification across Views (PReView), designed to learn predictive relationships between observations of the same identity captured by different camera platforms. The method employs lightweight predictors to map global representations from a source view to region-level latent targets extracted from a target view. This bidirectional prediction mechanism is jointly optimized with the standard ReID objective, encouraging the encoder to learn representations that facilitate both identity discrimination and cross-view prediction. The predictors and target encoders are utilized solely during training, while the standard retrieval pipeline remains unchanged during inference. Our method demonstrates promising retrieval performance for aerial-ground person re-identification.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Aerial-ground person re-identification, Cross-view predictive learning, Representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UXGjOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PReVIEW, a framework for aerial–ground person re-identification that introduces bidirectional cross-view feature prediction during training. Given an image from one platform, a platform-conditioned predictor uses its global representation and learnable region queries to predict regional features of a same-identity image from the other platform. An exponential-moving-average target encoder provides the prediction targets. The prediction objective is jointly optimized with identity classification and triplet losses. At inference, only the online encoder and BNNeck are retained, so the auxiliary components introduce no additional inference cost relative to the backbone baseline. Experiments on multiple aerial–ground benchmarks demonstrate substantial improvements over the reported Swin-B baseline.

Critical strengths

1.The paper clearly motivates the difficulty of learning identity representations across the substantial appearance gap between aerial and ground views. Its cross-view prediction objective provides a concrete way to exploit same-identity observations from both platforms, and the connection between the problem and the proposed training task is easy to follow.

2.The paper presents an extensive experimental evaluation across multiple aerial–ground person re-identification benchmarks and diverse retrieval protocols. It also includes ablation studies on prediction direction, prediction loss weight, and the number of target regions, together with qualitative retrieval examples. This broad coverage allows readers to assess the method under different evaluation settings and understand the influence of its key design choices, providing a substantial empirical basis for the paper.

Critical weaknesses

1.Although PReVIEW improves its own baseline, its advantages over strong existing methods remain insufficiently characterized. In Table 1, its CARGO cross-view mAP is 61.25, compared with 69.00 for ViSA and 70.53 for HiHR; it also does not lead any metric on AG-ReID.v2 in Table 5. A clearer comparison of backbones, pretraining, auxiliary supervision, and computational costs would help establish when PReVIEW offers an advantage. Its potential complementarity with stronger existing methods also remains unexplored.

2.Table 3 compares the classification-and-triplet baseline with variants of predictive supervision, but it omits a straightforward control: adding a direct feature-alignment loss between same-identity aerial and ground images. Without this comparison, it remains unclear whether a simpler additional alignment loss could yield similar improvements. This is particularly relevant because the introduction explicitly motivates PReVIEW through the limitations of direct alignment.

3.Figure 1 motivates the method using concrete identity cues, such as aerial head–shoulder contours and ground-view clothing textures or shoe details. However, the experiments do not examine whether PReVIEW actually retains and uses these cues more effectively than direct alignment. Figure 4 only shows whether the final retrieval is correct; it does not identify which cues explain the improvement. Thus, the claimed preservation of view-specific identity information lacks a concrete experimental example.

Questions for authors

1.Could you clarify the comparison conditions and computational costs relative to strong existing methods, and evaluate whether the proposed predictive supervision provides complementary benefits when incorporated into a stronger baseline?

2.Could you add a “baseline + direct cross-view feature alignment” control, using the same backbone, paired samples, and training settings, with appropriate loss-weight tuning, and compare it with PReVIEW?

3.Could you select a concrete cue from Figure 1 and compare PReVIEW with a direct-alignment baseline when that cue is visible versus selectively masked, to test whether the claimed advantage is associated with using that cue?

If these concerns are satisfactorily addressed, I will raise my Rating.

2 comments

W1+Q1: We thank the reviewer for recognizing the research motivation and experimental results of PReVIEW. We further clarify the advantages and applicable scope of the method relative to existing strong methods. First, under the CARGO AG cross-view protocol, PReVIEW achieves an mAP of 61.25%, which is lower than ViSA's 69.00% and HiHR's 70.53%. However, our main goal is not to comprehensively surpass existing methods by introducing a more complex retrieval architecture, but to explore whether cross-view latent prediction relationships can serve as complementary supervision to improve identity representation quality. Unlike some methods that adopt explicit view disentanglement, local relationship modeling, or generative feature completion, PReVIEW introduces an EMA target encoder and a prediction module only during training. At inference, it still uses the original identity encoder and BNNeck and does not add additional retrieval branches or feature dimensions.

Under the same Swin-B backbone and identity supervision setting, PReVIEW improves CARGO AG Rank-1/mAP from 56.88%/51.17% to 68.75%/61.25%, an increase of 11.87/10.08 percentage points, respectively. To examine applicability across different backbones, we apply PReVIEW to ViT-B/16, achieving 72.12%/65.30% and 63.13%/57.15% under the CARGO ALL and AG protocols, respectively. As a reference, the ViT baselines reported in existing literature are 61.54%/53.54% and 43.13%/40.11%, respectively. These results show that PReVIEW is not limited to the specific structure of Swin-B. Regarding computational overhead, we added training cost statistics under the same experimental environment (λ=0.2). Compared with the Swin-B baseline, PReVIEW increases total training time from 2 h 33 min 38 s to 3 h 15 min 40 s (+27.36%), average per-step time from 0.1889 s to 0.2404 s (+27.26%), and peak GPU memory from 9.48 GiB to 10.48 GiB (+10.53%).

These overheads occur only during training, and the inference structure and retrieval descriptor dimension remain unchanged. Therefore, we believe that the main advantage of PReVIEW is improving identity representation through predictive auxiliary supervision without adding inference modules. We also acknowledge that the additional benefit of combining with other stronger methods has not yet been directly verified and will be supplemented later.

W2+Q2: We thank the reviewer for suggesting a control experiment on direct cross-view feature alignment. We added a Direct Feature Alignment control experiment. Under the same Swin-B backbone, same-identity cross-platform pairing strategy, and training configuration, we retained the same EMA target encoder, stop-gradient mechanism, and bidirectional pairing as PReVIEW, but removed the learnable predictor and directly minimized the squared distance between the source-view online global representation and the target-view EMA global representation. Under the CARGO AG cross-view protocol, the standard Swin-B baseline achieves 56.88% Rank-1 and 51.17% mAP, Direct Alignment achieves 49.38%/43.28%, and PReVIEW achieves 68.75%/61.25%. Compared with direct alignment, PReVIEW improves Rank-1 by 19.37 percentage points and mAP by 17.97 percentage points, respectively. We additionally added a Global-to-Global Prediction control experiment. Under the AG protocol, it achieves 67.50%/60.65%. This result further indicates that under the current training configuration, learnable cross-view predictive supervision can obtain clear benefits, rather than achieving identity learning only by directly reducing the distance between the two view representations.

Based on the existing results, we believe that a preliminary conclusion can be drawn: under the current controlled experimental conditions, direct cross-view feature alignment cannot replace learnable latent predictive supervision, and learnable cross-view prediction provides significantly better performance than direct feature alignment.

W3+Q3: We thank the reviewer for this important suggestion. We agree that the retrieval examples in Figure 4 do not directly establish which identity cues contribute to the improvement. The view-specific cues illustrated in Figure 1 primarily motivate our predictive formulation rather than constitute direct evidence of cue preservation.

To further investigate this motivation, we conducted controlled comparisons under the same Swin-B backbone, training budget, random seed (42), and loss weight (λ=0.2). On the original CARGO AG protocol, Direct Alignment achieves 49.38% Rank-1 / 43.28% mAP, Global-to-Global Prediction achieves 67.50% / 60.65%, and PReVIEW achieves 68.75% / 61.25%. We acknowledge that these results do not isolate specific cues such as clothing textures, shoe details, or head-shoulder contours. However, we mainly learn a mapping and do not aim to delineate fine-grained specific cues.

Thank you for the detailed response and additional experiments. I consider my first and second concerns addressed. For the first concern, please incorporate the clarified scope, computational costs, and comparison limitations into the manuscript so that the core contribution is stated clearly and the claims remain proportionate to the evidence. Regarding the third concern, I would appreciate either additional experiments directly examining the preservation of view-specific, identity-relevant cues, or an explicit commitment to revise the corresponding mechanistic claims throughout the paper. The current control experiments demonstrate differences between training objectives but do not directly establish cue preservation. At this stage, I will maintain my current rating.

Reviewer_2jFfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an identity-conditioned cross-view latent prediction method, Predictive Re-identification across Views (PReView). The method combines an online Swin-B encoder, an EMA target encoder,a platform-conditioned MLP predictor, and a bidirectional aerial-to-ground/ground-to-aerial prediction loss.The prediction modules are used only during training, while the inference still uses the standard ReID encoder. Results are improved sustaintially on the authors’ matched Swin-B baseline and several benchmarks.

Critical strengths

Firstly, the main idea of PReVIEW is simple and useful. Instead of forcing aerial and ground features to be very similar, the model learns to predict the feature of one view from the other view. This is different from many existing alignment-based methods. Also, the predictor and the EMA encoder are only used during training, so there is no extra retrieval branch during testing. This makes the method more practical. Secondly, PReVIEW improves the authors’ Swin-B baseline on several datasets and protocols. This suggests that the prediction loss is not only useful on one specific dataset. The improvements are especially clear on CARGO A↔G and LAGPeR. So, although the paper does not fully explain why the method works, the results do show that the method is effective to some extent.

Critical weaknesses

The paper does not clearly show why the proposed mechanism works. There are no comparisons with direct feature alignment, global-to-global prediction, same-view prediction, or wrong-identity pairing. The paper also does not test what happens without EMA or stop-gradient. Because of this, it is still unclear whether the improvement really comes from the proposed cross-view global-to-region prediction, or just from adding another regression loss. The claim about preserving view-specific identity information is also not directly supported.The experimental evidence is still not strong enough. The paper does not report random seeds, results from multiple runs, standard deviations, or significance tests. Also, the main baseline is only Swin-B. The paper does not show whether PReVIEW also works with common ViT or TransReID backbones. In Table 3, the bidirectional setting is not always the best one. For example, G→A only has higher mAP than the bidirectional setting on G↔G. Therefore, the claims that bidirectional prediction is better and that the method is generally applicable may be a little too strong.

Questions for authors
  1. In Section 4.1, the paper says that five datasets are used, but six datasets are listed. Was CP2108 added as an extra experiment? Does the reported number of 15 protocols exclude CP2108?
  2. In Table 3, G→A only has higher G↔G mAP than the bidirectional setting, while A→G only reduces Rank-1 on some same-view protocols. Do the authors think that bidirectional prediction is generally better than unidirectional prediction?
  3. Please try to test PReVIEW with a ViT or TransReID backbone? If not, please explain why Swin-B was chosen and whether the method depends on the spatial feature structure of Swin-B?
Reviewer_51peOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PReVIEW, a cross-view predictive learning framework for aerial-ground person re-identification. The method predicts target-view regional representations from source-view global features using identity-conditioned cross-view pairs. The predictive objective is jointly optimized with conventional ReID losses, while the retrieval architecture remains unchanged during inference.

Critical strengths

The paper addresses an important challenge in aerial-ground person re-identification by exploring predictive learning as a complementary approach to conventional feature alignment. The motivation is reasonable, as aerial and ground observations often contain different identity-related cues that may not be effectively captured through direct alignment.

Critical weaknesses
  1. The methodological novelty appears incremental, as the framework builds on established predictive learning approaches such as BYOL and I-JEPA. The benefits of the proposed cross-view formulation over simpler predictive objectives are not sufficiently demonstrated.
  2. The proposed method does not consistently outperform existing state-of-the-art approaches. Although it achieves the best results on LAGPeR, it remains behind leading methods on CARGO A↔G, AG-ReID, and AG-ReID.v2, which limits its generalizability.
  3. The paper claims that predictive learning preserves view-specific identity information that may be lost through direct alignment. However, no direct quantitative analysis or controlled comparison is provided to validate this claim.
  4. The ablation studies do not fully justify several design choices.
  5. The predictor estimates target-view regional representations using only source-view global features and learned embeddings. It remains unclear how this formulation handles substantial variations in pose, occlusion, and visible body regions. The paper also lacks evidence that the predicted features retain meaningful regional information.
  6. The experiments do not report results across multiple random seeds. The additional training time and memory requirements of the prediction framework are also not analyzed.
  7. The paper contains some issues with writing and presentation. In particular, Lines 158–159 of the Related Work section contain an incomplete sentence. The number of evaluated datasets is also inconsistently stated.
Questions for authors
  1. How does PReVIEW compare with simpler predictive learning objectives adapted to same-identity cross-view pairs?
  2. Why does the method perform better on LAGPeR but remain less competitive under CARGO A↔G and on AG-ReID.v2? Which dataset characteristics explain these differences?
  3. Can the authors provide quantitative evidence that predictive learning preserves view-specific identity information better than direct feature alignment?
  4. What is the advantage of predicting regional representations instead of a single global target representation?
  5. How does the predictor preserve meaningful regional information when source and target images exhibit different poses and visibility patterns?
  6. Can the authors provide additional ablations examining the EMA target encoder, region queries, and platform embeddings?
  7. Can the authors report the additional training cost and performance variation across multiple random seeds?
1 comment

W1+Q1: We thank the reviewer for this concern. BYOL predicts representations across augmentations of the same image, while I-JEPA predicts target-region latents from context within one image. In contrast, PReVIEW learns identity-conditioned relations between distinct aerial and ground views, which can reveal different identity cues without reliable spatial correspondence. Our contribution is using cross-view latent prediction as complementary supervision alongside ID and triplet losses. On CARGO AG (Rank-1/mAP%), Swin-B obtains 56.88/51.17, direct alignment 49.38/43.28, global-to-global (G2G) prediction 67.50/60.65, and PReVIEW 68.75/61.25. Thus, learnable cross-view prediction is beneficial even with a global target; regional targets provide a further gain on AG.

W2+Q2: We acknowledge that PReVIEW is not SOTA on every protocol, including CARGO AG and AG-ReID.v2. Our goal is to improve representation learning through a training-only auxiliary objective, not to replace specialized retrieval architectures. PReVIEW improves the matched Swin-B baseline by 11.87/10.08 points on CARGO AG and achieves leading results on LAGPeR, without extra inference branches. These results support its usefulness across evaluated settings, but do not establish universal superiority. We will distinguish matched-baseline gains from absolute SOTA comparisons and avoid attributing dataset differences to unverified characteristics.

W3+Q3: We do not claim that ID/triplet learning or all alignment methods necessarily discard view-specific cues. Direct feature agreement and learnable cross-view prediction impose different supervision. On CARGO AG, alignment obtains 49.38/43.28 versus 68.75/61.25 for PReVIEW. Direct alignment tends to reduce the difference between source and target representations, while PReVIEW predicts the regional latent representation of another view through a learnable mapping and does not require the two representations to directly coincide in the feature space, providing another training approach for exploiting identity-related differential information across views.

W4+Q4: Region-level targets retain spatially differentiated supervision that global pooling removes, without requiring source-to-target part correspondence. Under the same Swin-B and λ=0.2, G2G reaches 67.50/60.65 on CARGO AG, while global-to-region reaches 68.75/61.25 (+1.25/+0.60 points); on AA, the respective results are 77.50/73.62 and 80.00/76.12. However, G2G has higher mAP on ALL and GG. These results show that cross-view latent prediction itself is an effective supervision mechanism, and the region-level target is a further refinement designed for spatial differences in the target view.

W5+Q5: PReVIEW does not explicitly estimate pose, recover invisible details or align corresponding body parts. A source global descriptor summarizes visible evidence; region queries and platform embeddings condition predictions of the other view's regional latent targets. The regions index target spatial locations, not predefined anatomical parts, so substantial viewpoint differences do not require one-to-one regional matches. Bidirectional A→G and G→A supervision lets both platforms act as sources and targets while jointly updating the shared identity encoder. This process occurs in the latent feature space, not the pixel reconstruction space, and therefore does not rely on precise local correspondence between the target image and the source image. Bidirectional prediction does not enable the model to reconstruct all invisible identity details, nor does it mean that the mapped features have precise semantic correspondence with the real local appearance. Its main role is to provide complementary cross-view learning constraints for the encoder through regional latent prediction in two directions. At the same time, identity classification and metric learning objectives jointly constrain the identity discriminative capability of the final retrieval representation.

W6+Q6: The reported experiments use seed 42; we will add runs with other seeds and report mean±standard deviation. For Swin-B versus PReVIEW (λ=0.2), training time is 2h33m38s versus 3h15m40s (+27.36%); per-step time is 0.1889s versus 0.2404s (+27.26%); and recorded peak GPU memory is 9.48 versus 10.48 GiB (+10.53%).

W7: We will correct the incomplete Related Work sentence (Lines 158–159) and the dataset count: the original submission evaluates six datasets including CARGO, LAGPeR, AG-ReID, AG-ReID.v2, G2APS-ReID and CP2108. We will also make the protocol counts consistent across the text and appendix.

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

Related papers

Discussion (0)

Sign in to comment.