CMNet : Cascading Moglow models for environment-aware motion synthesis

Abstract

We present CMNet, aimed at synthesizing human motions interacting with diverse sittable objects. Leveraging the MoGlow[1] architecture, our approach involves two distinct models - the Contact Generator and the Full-body Pose Generator. The former generates plausible contact positions for selected joints which correlates with chairs, while the latter synthesizes full-body motions based on these contacts, facilitated through a cascaded approach. Utilizing control signals derived from interaction between human motion and chair, our method ensures the synthesis of realistic and diverse human motions within various environmental settings. Experimental validation on the sorted CHAIRS dataset[2] demonstrates the efficacy of CMNet in generating diverse and realistic human interaction motions with chairs.