Low-Cost Neural Radiance Fields

TL;DR

This paper compares accelerated NeRF variants and explores extensions for low-resource settings, analyzing their performance and tradeoffs in reduced data and compute scenarios.

Contribution

It introduces modifications like depth supervision and architectural variants of HashNeRF, evaluating their effectiveness under constrained computational and data conditions.

Findings

None of the proposed extensions conclusively outperform baselines under iso-time evaluation.

Extensions transfer differently to constrained settings, informing future surface design.

Experimental results characterize tradeoffs in PSNR and training time for various architectures.

Abstract

Neural Radiance Fields (NeRF) achieve high-quality novel-view synthesis, but their long training times and reliance on dense input views limit accessibility. We present a comparative study of three accelerated NeRF variants - DS-NeRF, TensoRF, and HashNeRF and explore extensions targeted at the low-compute, low-data regime. First, we add a depth-supervision loss derived from COLMAP keypoints to TensoRF (TensoRF-DS) and evaluate it on the LLFF dataset under reduced view counts. Second, we ablate the feature-decoding MLP of TensoRF and study the effect of input downsampling on PSNR and runtime on the synthetic Lego scene. Third, we propose four architectural variants of the HashNeRF color and density networks, including residual and convolutional designs, and report PSNR/training-time tradeoffs under matched iteration budgets. Under iso-time evaluation, none of our extensions conclusively outperform the published baselines, but the experiments characterize which extensions transfer to constrained settings and surface design questions for future work.