This study proposes ViSP, a non-parametric 3D human body reconstruction framework that reconstructs a full-body mesh from a single frontal RGB image. Existing pixel-aligned implicit reconstruction methods can represent detailed human geometry without ...
This study proposes ViSP, a non-parametric 3D human body reconstruction framework that reconstructs a full-body mesh from a single frontal RGB image. Existing pixel-aligned implicit reconstruction methods can represent detailed human geometry without relying on parametric body templates, but their performance is often affected by unstable 2D–3D projection alignment and limited global structure representation. To address these limitations, ViSP introduces silhouette-based projection and ViT–PIFu feature fusion.
The silhouette-based projection module estimates scale, offset, and rotation from the human silhouette, thereby stabilizing the alignment between 2D image features and 3D query points without requiring ground-truth camera calibration. The feature fusion module combines global features extracted from a Vision Transformer with local features extracted from a PIFu-based Hourglass network, allowing the model to capture both whole-body proportions and local surface details.
Experiments on the AI Hub Korean full-body dataset show that ViSP improves 3D geometric reconstruction quality in terms of CD, P2S, mIoU, and NC compared with existing reconstruction methods. Additional experiments on the THuman 2.0 dataset also demonstrate stable reconstruction performance across diverse body shapes and poses. Ablation studies confirm that silhouette-based projection is a key component for improving 2D–3D projection alignment, while ViT-based global features further enhance structural consistency. In an application experiment, ViSP-based reconstructed meshes achieved lower prediction errors in body composition and body size estimation than meshes reconstructed by comparison models and a multiview 2D image-based baseline.
These results indicate that the proposed silhouette-based projection and ViT–PIFu feature fusion can improve non-parametric 3D human body reconstruction without depending on parametric templates or predefined camera calibration. The proposed framework provides a practical basis for 3D body digital twin applications such as body measurement, mobile healthcare, virtual fitting, and digital human generation.