Enhancing 3D Human Body Reconstruction: A Local Information-Based Approach with Improved Generalization and Accuracy

The 3D reconstruction of the human body from a single RGB or RGB-D image holds extensive application potential in domains such as virtual reality and human-computer interaction. Traditional methods, reliant on global information for image feature extraction, utilize implicit functions for model reconstruction. However, their generalizability is hindered by dataset constraints. Our study proposes a novel framework that incorporates directed distance signed fields, normal information, and raw human model data, enhancing the initial normal map prediction network. When tested on the THuman2.0 dataset, our method, which emphasizes local information, outperforms traditional global information-based approaches, showcasing enhanced reconstruction accuracy and broader applicability.