HMI-Bench is a benchmark for evaluating multimodal large language models on fine-grained human motion understanding. Existing human-video benchmarks often focus on coarse action or overall motion recognition, while HMI-Bench emphasizes detailed character movements, human-environment interactions, motion intentions, micro-expressions, attributes, and contextual interpretation.
HMI-Bench contains videos collected from MotionBench and Ego-Exo4D, together with… See the full description on the dataset page:
https://huggingface.co/datasets/GuoruiSong/HMI-Bench.
