ArtHOI: Realistic 4D Human-Object Interaction скачать в хорошем качестве

ArtHOI: Realistic 4D Human-Object Interaction

In this AI Research Roundup episode, Alex discusses the paper: 'ArtHOI: Articulated Human-Object Interaction Synthesis by 4D Reconstruction from Video Priors' ArtHOI introduces a framework for synthesizing physically plausible interactions between humans and articulated objects without requiring 3D supervision. The system addresses common issues like mesh penetration and contact loss by utilizing 4D geometric reasoning derived from monocular video priors. By leveraging video diffusion models such as KLing, the method recovers consistent 4D scenes from simple 2D video sequences. It employs flow-based part segmentation using CoTracker and Segment Anything to distinguish between dynamic components and static regions. This approach effectively resolves monocular ambiguity to ensure stable joint optimization of both human and object motion. Paper URL: https://arxiv.org/abs/2603.04338 #AI #MachineLearning #DeepLearning #ComputerVision #4DReconstruction #HumanObjectInteraction #VideoDiffusion #MotionSynthesis Resources: GitHub: https://github.com/Inso-13/ArtHOI