ReLU Fields: The Little Non-linearity That Could

Text‐to‐3D Shape Generation

“…Here Tr i is the transmittance, α i is the alpha compositing value, δ i is the length between neighboring samples on the ray and Ĉ(r) is the rendered color for ray r. Here we mention some notable works that are used by methods in this survey. Voxel grid NeRF models [SSC22;CXG*22;FYT*22;KRWM22] store learnable explicit/implicit features in a voxel grid structure for fast learning due to the smaller or outright removal of the MLP predicting density and color. InstantNGP [MESK22] uses multi-resolution hash tables for storing learnable features.…”

GO-Surf: Neural Feature Grid Optimization for Fast, High-Fidelity RGB-D Surface Reconstruction

“…Di-rectVoxGO [30] also improves NeRF's training time by two orders of magnitude by adopting an explicit density voxelgrid with post-activation interpolation -that enables to model sharp boundaries -and a feature voxel-grid with a shallow MLP for view-dependent appearance. Plenoxels [37] and ReLU Fields [8] take this idea further dropping the reliance on any neural networks. While Plenoxels [37] relies on a spherical harmonics basis, ReLU Fields [8] apply post-activation interpolation.…”

“…Plenoxels [37] and ReLU Fields [8] take this idea further dropping the reliance on any neural networks. While Plenoxels [37] relies on a spherical harmonics basis, ReLU Fields [8] apply post-activation interpolation. TensoRF [3] tackles the issue of memory footprint by modelling the volume field as a 4D tensor that is factorized into multiple compact low-rank tensor components.…”

“…Although MLPs have shown an extraordinary ability to represent radiance fields of complex scenes with a low memory footprint, this comes at the cost of very long training and inference times. Direct optimization of trilinearly interpolated features on multi-resolution feature grids has become a powerful alternative representation due to its fast convergence that leads to orders of magnitude faster training and inference times [13,8,37,41,3,30]. Their higher memory requirement has also been addressed by using sparse feature grids, such as octrees [37], or multiresolution hash tables [13].…”