Decoding Children's Gait Behavior

This paper frames fine-grained pediatric gait analysis from ordinary RGB video as a new human-action-recognition problem. It introduces a dataset of more than 1,100 high-frame-rate, 60 FPS sequences from 110 participants aged 3–17, paired with synchronized anonymized pose data. Each session uses a five-second walk-around task to capture multiple viewpoints. According to the supplied abstract, current gait foundation models and multimodal large language models struggle with subtle, irregular clinical patterns, while the authors provide a unified end-to-end baseline. Exact tasks, metrics, comparative results, access conditions, and clinical validation cannot be confirmed from the abstract alone.