Blender-NeRF: A Monocular Dynamic Human Body Explicit Reconstruction and Rendering Method

The Neural Radiance Field (NeRF) can implicitly represent non-rigid human body models from monocular videos. However, existing NeRF methods have not yet generated high-quality textured 3D human body models from monocular dy-namic human body videos and applied them to 3D graphics software. To address this issue, we propose Blender-NeRF, which consists of three modules: the NeRF dynamic human repre-sentation module, the textured 3D human body reconstruction module, and the Blender rigging and rendering module. Firstly, we utilize the Vid2Avatar method to obtain textureless 3D human body models and colored images from different viewpoints. Next, we propose a multi-view consistency depth computation method that involves back-projecting the textureless 3D human body models to different viewpoints for depth calculation. Finally, we introduce the Blender rigging and rendering module to handle rigging, action editing, and shading rendering. Adequate experiments demonstrate that Blender-NeRF can explicitly reconstruct textured 3D human body models for editing and rendering in 3D graphics software.