Skip to content

Latest commit

 

History

History
64 lines (50 loc) · 2.37 KB

DATASET_PREPARE.md

File metadata and controls

64 lines (50 loc) · 2.37 KB
Raw

IMC 2021 Dataset:

# Download:
cd /your/repo/path/DetectorFreeSfM/SfM_dataset
mkdir IMC2021 && cd IMC2021
wget https://www.cs.ubc.ca/research/kmyi_data/imc2021-public/imc-2021-test-gt-phototourism.tar.gz

tar -xzvf imc-2021-test-gt-phototourism.tar.gz

# Convert data format:
cd /your/repo/path/DetectorFreeSfM/SfM_dataset
sh scripts/parse_imc.sh

ETH3D Dataset:

# Download all undistorted images of training and test datasets, and GT scans of training dataset:

cd /your/repo/path/DetectorFreeSfM/SfM_dataset
mkdir ETH3D_source_data && cd ETH3D_source_data

wget https://www.eth3d.net/data/multi_view_training_dslr_undistorted.7z
wget https://www.eth3d.net/data/multi_view_training_dslr_scan_eval.7z

wget https://www.eth3d.net/data/multi_view_test_dslr_undistorted.7z

7z x multi_view_training_dslr_undistorted.7z && rm multi_view_training_dslr_undistorted.7z
7z x multi_view_training_dslr_scan_eval.7z && rm multi_view_training_dslr_scan_eval.7z
7z x multi_view_test_dslr_undistorted.7z && rm multi_view_test_dslr_undistorted.7z

# Now we get all 25 scenes. Then, convert data format:
cd /your/repo/path/DetectorFreeSfM/SfM_dataset

# For SfM dataset:
python tools/parse_data/parse_eth3d_dataset.py --triangulation_mode False --output_base_dir SfM_dataset/eth3d_dataset

# For Triangulation dataset:
python tools/parse_data/parse_eth3d_dataset.py --triangulation_mode True --output_base_dir SfM_dataset/eth3d_triangulation_dataset

TexturePoorSfM Dataset:

Download dataset from here and place it under SfM_dataset folder.

tar -xvf TexturePoorSfM_dataset.tar

Training Dataset:

Our multi-view matching refinement model is trained on MegaDepth dataset. If you don't need training, please skip this part.

Firstly, download MegaDepth depth maps and undistorted images following LoFTR's instruction, and place them as following structure:

repo_path/megadepth
    - Undistorted_SfM
    - phoenix
        - S6
            - zl548
                - MegaDepth_v1

Then, download the multiview matching training scene indices from here and unzip it:

tar -xvf multiview_matching_indices.tar

Finally, place the indices under megadepth folder: