3D Mesh Generation Model of Human Body Based on Monocular Image and 2D Pose

In this paper, we propose a model designed to generate a single human 3D mesh from monocular images and 2D human poses, even in highly occluded environments. The model employs a well-established 2D pose detector to capture the 2D pose information of the human body. Additionally, our proposed AFF module extracts robust features in occluded conditions. The model's regressor, leveraging both 2D and 3D joint features, allows it to focus on the target person's positional region while estimating depth information to distinguish the person from other objects. Ultimately, the model regresses the parameters of a parametric human body model. We validate the proposed method on public benchmark datasets, which qualitatively and quantitatively demonstrate its robustness in recovering 3D human body meshes in occluded environments.