SinHuman3D: Novel Multi-View Synthesis and 3D Reconstruction from a Single Human Image

This research introduces SinHuman3D, a framework designed for multi-view synthesis and three-dimensional reconstruction from a single human image, utilizing deep learning techniques in conjunction with Neural Radiance Fields (NeRF) technology. The framework integrates a global feature extraction module, a detailed feature extraction component, and a Transformer-based feature fusion approach, alongside the incorporation of prior knowledge from the SMPL human model. An effective feature extraction and fusion strategy is proposed for high-fidelity 3D human model reconstruction. Experimental validation on the Render People dataset demonstrates SinHuman3D's superior performance in 3D reconstruction quality, structural fidelity, and perceptual similarity, achieving a PSNR of 21.72, an SSIM of 0.87, and an LPIPS of 0.14, highlighting its effectiveness in human 3D reconstruction.