A NeRF benchmark suite that admits it’s not city-scale yet

What it does

UnboundedNeRFPytorch is a PyTorch benchmarking suite for neural radiance fields that target unbounded, real-world scenes. It trains and evaluates on Tanks & Temples, the Mip-NeRF-360 benchmark, and Waymo’s San Francisco Mission Bay using a unified pipeline. The authors report that their own approach outperforms NeRF++, Plenoxels, DVGO, and even Mip-NeRF-360 on the respective leaderboards, at least in PSNR.

The interesting bit

The authors renamed the project from LargeScaleNeRFPytorch because, as they note, it is not actually large enough to be city-scale—an unusual admission in a field prone to hype. They also trimmed the Waymo Block-NeRF dataset from 191 GB down to 19.1 GB by storing camera poses rather than raw ray directions, which is the kind of boring optimization that makes training less miserable.

Key highlights

• Claims PSNR 20.85 on Unbounded Tanks & Temples and 28.98 on the Mip-NeRF-360 benchmark, slightly edging the baselines in the README tables.
• Supports three major unbounded datasets: Tanks & Temples, Mip-NeRF-360, and Waymo’s San Francisco Mission Bay.
• Offers a config-driven workflow built around a single driver script, run_FourierGrid.py.
• Reduces the Waymo Block-NeRF dataset footprint from 191 GB to 19.1 GB by storing camera poses instead of raw ray directions.
• Explicitly tagged as a research project in progress, with a changelog that tracks the weekly deluge of new NeRF papers.

Caveats

• Still explicitly labeled a research project in progress.
• The README warns that custom scene reconstruction demands careful COLMAP parameter tuning to match reported quality.
• The published benchmark comparison focuses on PSNR; SSIM and LPIPS are supported via optional flags but not foregrounded in the main results table.

Verdict

Worth a look if you need a compact PyTorch baseline for unbounded NeRF research and can tolerate work-in-progress rough edges. Skip it if you want a turnkey, city-scale production pipeline; the authors freely admit this is not that.