Arxiv paper – GenHMR: Generative Human Mesh Recovery

In this episode, we discuss GenHMR: Generative Human Mesh Recovery by Muhammad Usama Saleem, Ekkasit Pinyoanuntapong, Pu Wang, Hongfei Xue, Srijan Das, Chen Chen. The paper introduces GenHMR, a novel generative framework for human mesh recovery (HMR) that addresses uncertainties in converting 2D images to 3D mesh. It employs a pose tokenizer and an image-conditional masked transformer to learn distributions of pose tokens, improving upon deterministic and probabilistic approaches. The model also includes a 2D pose-guided refinement technique and demonstrates superior performance compared to current methods.