Rectifying Geometry-Induced Similarity Distortions for Real-World Aerial-Ground Person Re-Identification

TL;DR

This paper addresses the core challenge in aerial-ground person re-identification: geometry-induced distortions that render the standard dot-product similarity in attention unreliable across extreme viewpoint changes. It introduces Geometry-Induced Query-Key Transformation (GIQT), a lightweight, low-rank module that explicitly rectifies the cross-view similarity space, paired with geometry-conditioned prompts to provide global priors aligned to camera geometry. The approach is model-agnostic and supports both metadata-rich and metadata-free settings via a vision-based geometry predictor. Across four benchmarks, the proposed GeoReID framework consistently improves robustness under unseen geometric conditions with minimal computational overhead, demonstrating the critical importance of explicit geometry-aware similarity modeling for real-world AG-ReID tasks.

Abstract

Aerial-ground person re-identification (AG-ReID) is fundamentally challenged by extreme viewpoint and distance discrepancies between aerial and ground cameras, which induce severe geometric distortions and invalidate the assumption of a shared similarity space across views. Existing methods primarily rely on geometry-aware feature learning or appearance-conditioned prompting, while implicitly assuming that the geometry-invariant dot-product similarity used in attention mechanisms remains reliable under large viewpoint and scale variations. We argue that this assumption does not hold. Extreme camera geometry systematically distorts the query-key similarity space and degrades attention-based matching, even when feature representations are partially aligned. To address this issue, we introduce Geometry-Induced Query-Key Transformation (GIQT), a lightweight low-rank module that explicitly rectifies the similarity space by conditioning query-key interactions on camera geometry. Rather than modifying feature representations or the attention formulation itself, GIQT adapts the similarity computation to compensate for dominant geometry-induced anisotropic distortions. Building on this local similarity rectification, we further incorporate a geometry-conditioned prompt generation mechanism that provides global, view-adaptive representation priors derived directly from camera geometry. Experiments on four aerial-ground person re-identification benchmarks demonstrate that the proposed framework consistently improves robustness under extreme and previously unseen geometric conditions, while introducing minimal computational overhead compared to state-of-the-art methods.

Figures (11)

• Figure 1: Comparison between the performance degradation of the well known SeCap method as baseline and our proposal, across altitude and angle bins. The y-axis povide the $\Delta$ mAP = (Bin mAP - Overall SeCap A→G mAP). Values above zero indicate better-than-overall performance and values below zero indicate perfomance degradation. Overall, it is evident that the proposed method keeps superior performance in extreme geometric regimes.
• Figure 2: Overview of the proposed geometry conditioned similarity alignment framework for aerial-ground person ReID. Camera geometry guides both global prompt adaptation and local similarity rectification via the proposed Geometry Induced Query Key Transformation (GIQT).
• Figure 3: Vision only multi task geometry prediction architecture. A shared ResNet-50 encoder extracts visual features from RGB images, followed by task specific heads for altitude and distance regression and viewing-angle classification.
• Figure 4: Singular value spectrum and cumulative energy distribution of cross view feature covariance difference, showing highly anisotropic geometry induced distortion.
• Figure 5: Performance sensitivity with respect to GIQT rank across datasets. Low rank correction (8--16) yields optimal performance.