PD-NeRF: a general pseudo-depth supervision method for neural radiance fields
Gu J M, Jiang M C, Lu X Y, et al
Sci China Inf Sci, 2025, 68(8): 184101
Synthesizing novel views of a scene and reconstructing a 3D scene from a sparse set of captured images has been a long-standing challenge in computer vision. Neural radiance fields (NeRF) is a seminal work that introduced a breakthrough approach to 3D reconstruction and novel view synthesis. NeRF differs from traditional 3D reconstruction methods, which represent scenes using explicit structures such as point clouds, grids, and voxels. NeRF sampling points along each ray, determine the 3D location x = (x, y, z) of each sampling point and the 2D viewing direction of the ray. These 5D vectors are then fed into a neural network to obtain the color c = (r, g, b) and volume density of the sampling point. In other words, NeRF constructs a field parameterized by an multilayer perceptron (MLP) neural network to reconstruct the scene and continuously optimize parameters. However, the traditional NeRF method requires nearly a week to train a single scene. In addition, rendering speed is slow, generating a single image takes several minutes, and the resulting scene reconstruction often lacks fine detail.