Abstract Neural Radiance Fields (NeRFs) have achieved great success in the past few
years. However, most current methods still require intensive resources due to ray marching-
based rendering. To construct urban-level radiance fields efficiently, we design Deformable
Neural Mesh Primitive (DNMP), and propose to parameterize the entire scene with such
primitives. The DNMP is a flexible and compact neural variant of classic mesh
representation, which enjoys both the efficiency of rasterization-based rendering and the …

The overall framework is implemented using Py-Torch [11]. The differentiable rasterization is
implemented based on PyTorch3D [12]. The network Fθ is composed of 8 layers with width
256 for opacity prediction and additional 2 layers for view-dependent radiance value
prediction. In our lightweight version, the layer number and width of the MLPs for opacity
prediction are reduced to 2 and 64, respectively. We use positional encoding with frequency
L= 4 to encode the view-dependent factors. As mentioned in the manuscript, we use Mip …