基于锥形追踪和网络分解的NeRF三维重建方法

NeRF 3D Reconstruction Method Based on Cone Tracking and Network Decomposition

在计算机视觉领域,神经辐射场(NeRF)是以空间坐标或者时间、相机位姿等其他维度作为输入,通过多层感知机(MLP)网络模拟目标函数,生成颜色、深度等目标标量的过程。NeRF的应用包括对三维场景进行高质量的重建,而其在处理不同分辨率的场景时会产生过度模糊或者伪影的渲染效果,且存在训练耗时较长的问题。为了解决上述问题,提出基于锥形追踪和网络分解的NeRF三维重建方法。使用锥形追踪的方法,为每个像素投射一个圆锥体,并将投射的圆锥体切割成一系列的圆锥台,沿着该圆锥体进行特征化,通过高效渲染抗锯齿的圆锥台来降低模糊或者伪影效果。为了缩短训练时间,使用网络分解的方法,将原始NeRF接收5维数据的神经网络分解为两个网络,有效地缩短训练时间。实验结果表明,在NeRF_Synthetic、LLFF和Multiresolution数据集中,相比于NeRF、F 2-NeRF等方法,所提方法的峰值信噪比(PSNR)提升了14.4%~24.6%,能够重建出更丰富的细节特征,视觉效果更好,且训练时间大幅降低。

In computer vision,Neural Radiance Fields(NeRF)define processes that use spatial coordinates or other dimensions,such as time and camera pose,as input and simulate the objective function through a Multi-Layer Perceptron(MLP)network to generate the target scalar(color and depth).NeRF reconstructs 3D scenes well but blurs or distorts different resolutions and trains them slowly.To solve these two issues,this study proposes a NeRF 3D reconstruction method based on cone tracking and network decomposition.First,the cone-tracking method is used to project a cone for each pixel;the projected cone is cut into a series of cones,characterized along the cone,and the blur or artifact effect is reduced by efficiently rendering the anti-aliasing cone.To shorten the training time,the neural network of the original NeRF receiving five-dimensional data is decomposed into two networks using the network decomposition method,which effectively shortens the training time.Experimental results show that the proposed method improves the Peak Signal-to-Noise Ratio(PSNR)by 14.4%-24.6%compared with NeRF,F 2-NeRF,and other algorithms in NeRF_Synthetic,LLFF,and Multiresolution datasets.The training time is also reduced,which allows the reconstruction of richer detailed features,better visual effects,and faster training speed.