DATASETS.md

Datasets

DTU (small baseline)

• Download the preprocessed DTU training data dtu_training.rar. Also download Depth_raw.zip if would like to evaluate the depth accuracy, otherwise no depth is needed for training.

• Extract Cameras/ and Rectified/ from the above downloaded dtu_training.rar, and optionally extract Depths from the Depth_raw.zip. Link the folders to DTU, which should have the following structure:

DTU
├── Cameras
└── Rectified

DTU (large baseline)

• Please refer to RegNeRF for the downloading of the DTU dataset.
• The folder structure:

DTURaw/
├── Calibration
├── idrmasks
└── Rectified

RealEstate10K

• Please refer to AttnRend for the downloading of the RealEstate10K dataset.
• The folder structure (data_download contains video frames, and RealEstate10K contains camera poses):

realestate_full
├── data_download
│   ├── test
│   └── train
└── RealEstate10K
    ├── test
    └── train

• The full RealEstate10K dataset is very large, which can be challenging to download. We use a subset provided by AttnRend for ablation experiments in our paper.
• The folder structure of the subset:

realestate_subset
├── data_download
│   └── realestate
│       ├── test
│       └── train
└── poses
    └── realestate
        ├── test.mat
        └── train.mat

LLFF

• Please refer to IBRNet for the downloading of the mixed training datasets.
• Download the LLFF test data with:

gdown https://drive.google.com/uc?id=16VnMcF1KJYxN9QId6TClMsZRahHNMW5g
unzip nerf_llff_data.zip

• The folder structure:

mixdata1/
├── google_scanned_objects
├── ibrnet_collected_1
├── ibrnet_collected_2
├── nerf_llff_data
├── nerf_synthetic
├── RealEstate10K-subset
├── real_iconic_noface
└── spaces_dataset

Mip-NeRF 360 dataset

• Download the dataset with

wget http://storage.googleapis.com/gresearch/refraw360/360_v2.zip

• The folder structure:

mipnerf360/
├── bicycle
├── bonsai
├── counter
├── flowers.txt
├── garden
├── kitchen
├── room
├── stump
└── treehill.txt