Skip to content

ReadMe.md

Latest commit

 

History

History
124 lines (115 loc) · 3.65 KB

ReadMe.md

File metadata and controls

124 lines (115 loc) · 3.65 KB
Raw

Data

Pretrained models

  • SAM2.1 : create the folder and download your desired checkpoint from the official repo. e.g.:
cd /<ovo_abs_path>/data/input/
mkdir sam_ckpts && cd sam_ckpts
wget https://dl.fbaipublicfiles.com/segment_anything_2/092824/sam2.1_hiera_large.pt
  • CLIPs merging weights predictor: dowload its weights from this link, and save it in /<ovo_abs_path>/data/input/weights_predictor/base/.

Datasets

Replica

Following OpenNeRF, we use Replica dataset trajectories processed for NICE-SLAM, and their computed semantic GT. We provide the GT inside ./data/input/replica_gt_semantics .

The expected dataset structure is:

/<ovo_path>/data/input/Datasets/Replica/
     -> semantic_gt/
        -> office0.txt
        .
        .
        .
        -> room2.txt
     -> office0/
         -> results/
         -> traj.txt
     -> office0_mesh.ply
     .
     .
     .    
     -> scene0704_01/
     -> room2/
     -> room2_mesh.ply

To download the data and link the semantic GT (you can also move or copy), run in the terminal:

cd <ovo_abs_path>
mkdir -p ./data/input/Datasets && cd ./data/input/Datasets/
wget https://cvg-data.inf.ethz.ch/nice-slam/data/Replica.zip && unzip Replica.zip
rm Replica.zip
cd Replica
ln -s /<ovo_abs_path>/data/input/replica_semantic_gt/ semantic_gt

ScanNet

The expected dataset structure is:

/<ovo_path>/data/input/Datasets/ScanNet/
     -> semantic_gt/
        -> scene0011_00.txt
        .
        .
        .
        -> scene0704_01.txt
     -> scannet200_gt/
        -> scene0011_00.txt
        .
        .
        .
        -> scene0704_01.txt
     -> scene0011_00/
         -> color/
         -> depth/
         -> pose/
         -> intrinsic/
         -> scene0011_00_vh_clean_2.labels.ply
     .
     .
     .    
     -> scene0704_01/
  • Follow the official repository https://github.com/ScanNet/ScanNet instructions to obtain a download link. Then download and decode the validation split to obtain two folders like:
/<ScanNet_data_path>/ 
 -> scans/
     -> scene0011_00/
     .
     .
     .
     -> scene0704_01/
 -> data/val/
     -> scene0011_00/
         -> color/
         -> depth/
         -> pose/
         -> intrinsic/
     .
     .
     .    
     -> scene0704_01/
  • Then, link GT point-clouds to data folder, and extract ScanNet20 semantic gt in data/semantic_gt running:
cd /<ovo_path>/
conda activate ovo
python scripts/scannet_preprocess.py --data_path /<ScanNet_data_path>/  --link_pcds

  • (Optional) Extract ScanNet200 semantic gt:

    • Follow instructions in /<ScanNet_repo_path>/BenchmarkScripts/ScanNet200/ReadMe.md to compute ScanNet200 labels. You should execute something like: 
      cd /<ScanNet_repo_path>/BenchmarkScripts/ScanNet200/
create -n scannet200 python=3.8
conda activate scannet200
pip install -r requirements.txt
python preprocess_scannet200.py --dataset_root /<ScanNet_data_path>/scans --output_root  /<ScanNet_data_path>/scannet200 --label_map_file /<ScanNet_repo_path>/Tasks/Benchmark/
      preprocess_scannet200.py by default preprocess all scenes in scans. If you want to preprocess only the validation split, you can add a return after lines 31 and 37 of preprocess_scannet200.py .
    • Then, run the following script to extract scannet200 labels into txts. 
      cd /<ovo_path>/
conda activate ovo
python scripts/scannet_preprocess.py --data_path /<ScanNet_data_path>/  --scannet200

  • Finaly move or link data folder to /<ovo_path/:

    ln -s /<ScanNet_data_path>/data/val/ /<ovo_path/data/input/Datasets/ScanNet