Three-Dimensional Human Body Reconstruction and Novel Views Synthesis Based on Bidirectional Constraints

For various applications such as virtual reality and digital entertainment, three-dimensional human body reconstruction and synthesis of novel views are crucial. Previous works have mostly employed unidirectional constraint deformation fields to transform the human body from observation space to canonical space. And they have also often overlooked the optimization of fine-grained image features when synthesizing novel views of the human body. Therefore, this paper proposes a new framework called HumanBC, which primarily consists of a bidirectional constraint neural offset deformation field and a multiresolution Gaussians mesh module. The former ensures the accuracy of the deformation field by employing bidirectional alignment between the observation and canonical spaces, thereby improving the reconstruction quality of the 3D human body model. The latter effectively reduces the artifacts in novel view synthesis by blending the fine-grained features of the keyframe images with the geometric features of the human body model in the canonical space. It demonstrated superior performance on a human body dataset created for this study.