Core Concepts
HOI-Diff enables realistic 3D human-object interactions driven by textual prompts through modular design and affordance prediction.
Abstract
The content introduces HOI-Diff, a model for generating realistic 3D human-object interactions based on text prompts. It decomposes the task into simpler sub-tasks, including a dual-branch diffusion model (HOI-DM) for motion generation and an affordance prediction diffusion model (APDM) for contact area estimation. The approach is evaluated on BEHAVE and OMOMO datasets, showcasing superior performance in generating diverse and coherent interactions.
Directory:
Introduction to HOI-Diff
Modular design for 3D HOIs synthesis.
Core Components
Dual-branch diffusion model (HOI-DM).
Affordance prediction diffusion model (APDM).
Evaluation on BEHAVE and OMOMO datasets.
Performance metrics: FID, R-Precision, Diversity, Contact Distance.
Ablation Studies
Impact of different components on performance.
Annotation Process for BEHAVE Dataset
Stats
HOI-Diffは、テキストプロンプトに基づいてリアルな3D人物オブジェクト相互作用を実現します。