PRADA: Prompt-driven Dual Alignment with Actor-Critic Rewards for Aerial-Ground Person Re-Identification

ICLR 2026 Conference Submission15378 Authors

Keywords: Aerial-ground person re-identification (ReID), Prompt-drivenDual Alignment (PDA), Part-level Actor-Critic Reward Engine(PARE), Dynamic Shift-Shuffle Grouping (SSG)

Abstract:

Aerial-Ground person Re-IDentification (AG-ReID) aims to match individuals captured from aerial drones and ground surveillance cameras, posing unique challenges due to severe viewpoint variations, scale discrepancies, and heterogeneous resolutions.
To address these challenges, we propose PRADA (Prompt-driven Dual Alignment with Actor-Critic Rewards), a unified framework that combines Prompt-driven Dual Alignment (PDA) and Part-level Actor-Critic Reward Engine (PARE) for robust cross-view representation learning.
Specifically, PDA enforces cross-platform identity consistency while preserving intra-platform specificity, and PARE dynamically emphasizes discriminative local features by optimizing complementary classification and confidence rewards under an Actor-Critic paradigm.
Additionally, a Cross-Platform Multi-Positive Alignment loss further aligns identity features across aerial and ground domains.
Extensive experiments on benchmark AG-ReID datasets demonstrate that PRADA outperforms state-of-the-art methods, validating the effectiveness of integrating prompt-guided alignment with reinforcement-based part-level supervision.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 15378