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,F2-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.
国家科技期刊平台
NSTL
万方数据
