3DPMesh: An enhanced and novel approach for the reconstruction of 3D human meshes from a single 2D image

Abstract

Currently, virtual reality, the gaming industry, cloth parsing, surveillance and security systems, and avatars have gained much interest in computer vision. To fulfil the growing demand in this field, an accurate 3D human mesh plays an important role. We proposed a novel approach, 3DPMesh, to reconstruct more accurate 3D meshes for humans from a single 2D image. 3DPMesh stands for 3DPose estimation, 3DPose enhancement and Mesh articulation. We achieve enhanced results of 3D meshes with the ultimate fusion of these three modules. The depth predictor regresses the 3D keypoints of the human body, whereas the pose alignment enhances the accuracy of the predicted 3D joint locations. Finally, the mesh articulation module articulates the parametric mesh so that it accurately matches the 3D pose of a person. The proposed method offers more accurate 3D human meshes and better alignment to the content of an input image when compared to the state-of-the-art methods on the challenging UP-3D, Human3.6M, and 3DPW datasets. Additionally, there is no need for a training phase in the proposed model, therefore, it does not require an extensive collection of datasets to train the model, so it saves a lot of computational time and ultimately generates a 3D mesh in less inference time. The experimental results and analysis of the proposed method indicate that it advances the mesh fitting state-of-the-art methods at any scale variation.