基于深度优化NeRF空间失真的三维重建方法研究

Research on 3D Reconstruction Method Based on Deep Optimization of NeRF Spatial Distortion

摘要: 在现有神经辐射场隐函数表达能力的限制下，针对其光线追踪终止深度的不确定性，深度均方差无法收敛，进而导致重建结果空间失真、远处渲染效果差的问题，提出了基于深度信息优化空间失真的神经辐射场。首先使用深度信息引导自适应采样方法：通过图像分割和Canny算法获取物体深度并解析边缘信息，据此优化采样点数量和权重分配；然后融合由深度损失Ld上的高斯负对数似然(GNLL)项、轮廓损失和颜色损失作为网络的损失函数，最小化预测误差；最后提出了一种高斯平滑的周期编码方式，降低噪声对位置信息的干扰。实验结果表明，本文融合的深度信息在三维重建中得以充分利用，保证渲染质量的同时降低空间失真，以增强空间逼真感。

Abstract: Under the limitations of the implicit function expression ability of existing Neural Radiance Fields (NeRF), the Research on 3D reconstruction method based on deep optimization of NeRF spatial distortion is proposed to address the uncertainty of ray tracing termination depth and the inability of depth mean square error to converge, resulting in spatial distortion of reconstruction results and poor rendering effects at a distance. Firstly, use depth information to guide adaptive sampling methods: obtain object depth and parse edge information through image segmentation and Canny algorithm, and optimize the number of sampling points and weight allocation accordingly; then, the Gaussian Negative Logarithmic Likelihood (GNLL) term that combines the mean square error of light color and depth information is used as the loss function of the network to minimize prediction error; finally, a Gaussian smoothing periodic encoding method was proposed to reduce the interference of noise on position information. The experimental results show that the depth information fused in this paper is fully utilized in 3D reconstruction, ensuring rendering quality while reducing spatial distortion to enhance spatial realism.