Learning What to Generate: A Reinforcement Learning-based Closed-Loop Augmentation Framework for Person Re-identification

Person re-identification (ReID) models are sensitive to long-tail nuisances (e.g., rare viewpoints, occlusions, complex backgrounds), yet current generative augmentation is largely open-loop: prompts/conditions are sampled heuristically without verifying whether the synthesized samples improve ReID discriminability. We introduce ReasonAug, a closed-loop framework that learns an image-conditioned instruction policy for a frozen generator, turning augmentation into a sequential decision problem over instruction tokens. A Semantic Reasoning Agent (SRA) performs hierarchical planning from global semantics to identity-critical local cues, producing structured edit instructions whose utility is verified by downstream ReID feedback. To make closed-loop optimization reliable, we propose Metric-Aligned Gated Reward (MAGR), which converts metric-learning objectives into a dense reward while gating task shaping by identity preservation to prevent reward hacking, and Structure-Aware Entropy (SAE), which allocates exploration per token to lock identity-critical cues while diversifying nuisance factors. Experiments on Market-1501 and MSMT17 demonstrate state-of-the-art performance, confirming that closing the augmentation loop and learning what to generate yield more discriminative training data than open-loop alternatives.

Person re-identification systems try to recognize the same person seen by different cameras—an ability used in applications like searching surveillance footage or analyzing customer flow in stores. These systems learn from photo collections, but real-world data is uneven: some viewpoints, outfits, and situations (such as side views, crowded scenes, or partially blocked figures) appear far less often than others, which makes the system unreliable in those cases. A common workaround is to use AI image generators to create extra training photos, but today this is done in an "open-loop" way: the system blindly produces images from generic prompts without checking whether they actually help. We introduce ReasonAug, which closes this loop. Instead of guessing what to generate, an AI agent learns, by trial and error with feedback, how to write better photo-editing instructions tailored to each person. Two ideas keep this process honest: a reward design that only rewards diversity when the person's identity is clearly preserved, and an exploration scheme that lets the agent freely vary background or pose while protecting identity-defining details like clothing color or accessories. On standard benchmarks (Market-1501 and MSMT17), ReasonAug produces noticeably more useful training images than previous methods, especially for the rare situations where re-identification typically fails. The work suggests that, with today's strong image generators, the key is no longer how to generate images but what to generate—and that question can be learned automatically.