Mip-NeRF 360 -- Bonsai (NO geometry GT)

mipnerf360_bonsai · Mip-NeRF 360 · planned — not present locally yet

Ground-truth mesh

Not present locally. See the acquisition plan below.

Why this sample is in the benchmark

Registered as an explicit NEGATIVE example: it has no ground-truth mesh, so it can only be used for appearance metrics. Including it in the catalog makes the "is this valid geometry GT?" question answerable from the registry itself (gt_mesh: none).

Use only for PSNR/SSIM/LPIPS and robustness, never for Chamfer/F-score.

Ground truth

GT type

none

licence

CC BY 4.0

source

Mip-NeRF 360

builder

- {'scene': 'bonsai'}

views

200 × 1080px (orbit, fov 60°)

lighting

provided

Acquisition plan

kind script, automatic False, estimated size 4000 MiB

1. Download the 360_v2 archive (~4 GB gzipped)

   bash -c 'curl -L -o 360_v2.zip http://storage.googleapis.com/gresearch/refraw360/360_v2.zip && unzip -q 360_v2.zip'

2. Auto-download via nerfbaselines

   pip install nerfbaselines && python -c "import nerfbaselines; nerfbaselines.download_dataset('mipnerf360')"

3. Project page
   https://jonbarron.info/mipnerf360/

Mip-NeRF 360: 9 unbounded indoor/outdoor scenes. IMPORTANT CAVEAT: this dataset ships NO ground-truth mesh -- it was collected for novel-view synthesis, not geometry. Use it for photo-realistic appearance / robustness checks, and pair it with Replica or Tanks & Temples for geometry. A COLMAP point cloud is provided, which is reconstruction-derived and therefore NOT valid ground truth.

Comparisons

No reconstruction compared yet. Run geombench compare mipnerf360_bonsai --mock or point it at a real 3DGS output.