Abstract
Human-object interaction (HOI) and human-scene interaction (HSI) are crucialfor human-centric scene understanding applications in Embodied ArtificialIntelligence (EAI), robotics, and augmented reality (AR). A common limitationfaced in these research areas is the data scarcity problem: insufficientlabeled human-scene object pairs on the input images, and limited interactioncomplexity and granularity between them. Recent HOI and HSI methods haveaddressed this issue by generating dynamic interactions with rigid objects. Butmore complex dynamic interactions such as a human rider pedaling an articulatedbicycle have been unexplored. To address this limitation, and to enableresearch on complex dynamic human-articulated object interactions, in thispaper we propose a method to generate simulated 3D dynamic cyclist assets andinteractions. We designed a methodology for creating a new part-basedmulti-view articulated synthetic 3D bicycle dataset that we call 3DArticBikesthat can be used to train NeRF and 3DGS-based 3D reconstruction methods. Wethen propose a 3DGS-based parametric bicycle composition model to assemble8-DoF pose-controllable 3D bicycles. Finally, using dynamic information fromcyclist videos, we build a complete synthetic dynamic 3D cyclist (riderpedaling a bicycle) by re-posing a selectable synthetic 3D person whileautomatically placing the rider onto one of our new articulated 3D bicyclesusing a proposed 3D Keypoint optimization-based Inverse Kinematics poserefinement. We present both, qualitative and quantitative results where wecompare our generated cyclists against those from a recent stablediffusion-based method.