基于神经辐射场的三维服装建模技术

3D garment modeling technology based on neural radiance field

针对三维服装建模技术存在的难操作、硬件要求高、质量和稳健性不佳等问题,提出一种改进的基于神经辐射场(NeRF)的快速三维服装建模技术。以简单的服装多视角图为输入,采用增量式运动恢复技术求取多视角图像的位姿,将服装图像与位姿信息送入改进的NeRF模型进行加速训练后,得到高质量的服装三维模型。为提高三维建模速度,采用体素网格显式地表示NeRF,从而缩短原始NeRF所需的训练时间,并通过三线性插值、从粗到细的网格优化技巧优化训练过程。基于标准3D服装建模数据集的试验结果显示,改进NeRF技术的建模效果良好,训练22 min后峰值信噪比(PSNR)可达30左右,可以满足不同种类服装的建模要求,相比原始NeRF建模具有一定的优势。

A rapid 3D modeling technique for garments based on improved neural radiance field(NeRF)is proposed to address the problems of difficult operation,high hardware requirements,and poor quality and robustness of existing 3D modeling techniques for garments.With a simple multi-view image of the garment as input,the incremental motion recovery technique is used to obtain the poses of the multi-view image.A high-quality 3D model of the garment is obtained after feeding the garment image and the poses information into the improved NeRF model.To improve the 3D modeling speed,a voxel grid is used to represent NeRF,thus reducing the training time required for the original NeRF.The training process is greatly optimized by trilinear interpolation and coarse-to-fine grid optimization techniques.The experimental results on the standard 3D garment reconstruction-based dataset show that the improved NeRF technique is effective in modeling,and the peak signal-to-noise ratio(PSNR)can reach about 30 after 22 minutes of training,which can meet the modeling requirements for different kinds of garments,and has an advantage over the original version of NeRF modeling.Compared with the original NeRF modeling,it has some advantages.