The arXiv announcement introduces HumanPhys, a benchmark designed to assess AI's physical perception in real-world settings with a focus on people. The benchmark comes with a training corpus produced by the same pipeline, suggesting a coordinated effort to both evaluate and improve models.
Each question in HumanPhys targets a physically grounded quantity related to people, with depth explicitly mentioned in the abstract. The broader goal is to push visual reasoning beyond object recognition toward understanding physical properties of human scenes.
The abstract is brief, so details on the full range of quantities or evaluation methodology are not yet available. Still, the benchmark appears aimed at closing the gap between controlled visual tasks and the messy physical world humans inhabit.