Automatic 3D character reconstruction from frontal and lateral monocular 2D RGB views

Abstract

We present an automatic animatable 3D character reconstruction method from frontal and lateral RGB pictures of the person to reconstruct. We propose a method where, after obtaining a 3D reconstruction from a frontal view, we adapt the template of the Skinned Multi-Person Linear Model (SMPL) to it. This frontal modified SMPL is then used in a second step as an input to reconstruct the character from the lateral view, thus obtaining a correct person reconstruction in all dimensions. Our method enables the creation of 3D animatable characters focusing on two main aspects: shape and texture. The obtained results show that the proposed method achieves an accurate automatic 3D character reconstruction from only two views that enhances current methods while offering a reproducible work-flow.