Compact Half-Side Body Linear Blending Skinning Weights for Human NeRF Representation

Human body reconstruction using Neural Radiance Fields (NeRF) based on monocular video sequences faces challenges due to limited field of view, leading to insufficient coverage of motion and appearance information, which negatively impacts motion editability and image rendering quality. To address these issues, we propose a novel compact half-side weight human NeRF representation. First, a canonical human pose model is constructed based on NeRF; then, learnable half-side linear blend skinning weights are introduced to leverage human body symmetry to enhance motion expressiveness, enabling precise transformation from specific poses to canonical poses. Finally, a structured weight decay and weight-based sampling strategy is employed to improve volumetric rendering efficiency and quality. Experimental results show that the proposed method exhibits strong editability for unseen motions on the ZJU-MoCap dataset, achieving SSIM, PSNR, and LPIPS scores of 0.957, 29.16, and 0.028, respectively, outperforming the compared methods in all metrics.