Message from the Best Paper Award Committee

Summary: This paper presents a method that infers object properties from human interaction motion alone, showing the strong correlation between motion and interacting objects. Results indicate that related object latent properties can be inferred solely from 3D skeleton sequences, enabling new synthesis possibilities for motions involving human interaction.

Summary: This paper proposes an efficient framework for human pose estimation by customizing backbone network design and compressing transposed convolutions to reduce computational cost while maintaining accuracy. Experimental results demonstrate the excellent performance of the framework on MPII and COCO datasets.