iMatching Dataset Overview

Dataset

TartanAir

• Source: TartanAir
• Download Tool: tartanair_tools
• Download Command:

  python download_training.py --output-dir OUTPUTDIR --rgb --depth --only-left
  

• Data Structure:

  $DATASET_ROOT/
  └── tartanair/
      ├── abandonedfactory_night/
      |   ├── Easy/
      |   └── ...
      │   └── Hard/
      │       └── ...
      └── ...
  

• Notes:
  • Only <ENVIRONMENT>/<DIFFICULTY>/<image|depth>_left.zip files are required.
  • After extraction, remove the duplicated <ENVIRONMENT> directory level.

ETH3D SLAM

• Source: ETH3D SLAM
• Download Method: The dataset will be automatically downloaded by the datamodule.

Pre‑trained Weights

• Download Command:

  pip install gdown
  mkdir pretrained
  cd pretrained/
  # CAPS
  gdown 1UVjtuhTDmlvvVuUlEq_M5oJVImQl6z1f
  # p2p
  sh ../ext/patch2pix/pretrained/download.sh
  # aspan
  gdown 1eavM9dTkw9nbc-JqlVVfGPU5UvTTfc6k
  tar -xvf weights_aspanformer.tar
  cd ..
  

Training

Train on TartanAir

• CAPS:

  scene=abandonedfactory
  d=Easy
  python ./train.py data_root=./data/datasets datamodule.include="$scene\_$d"
  

• Patch2Pix:

  scene=abandonedfactory
  d=Easy
  python ./train.py --config-name p2p-train-tartanair data_root=./data/datasets trainer.max_epochs=2 datamodule.include="$scene\_$d"
  

• AspanFormer:

  scene=abandonedfactory
  d=Easy
  python ./train.py --config-name aspan-train-tartanair data_root=./data/datasets trainer.max_epochs=2 datamodule.include="$scene\_$d"
  

• DKM:

  scene=abandonedfactory
  d=Easy
  python ./train.py --config-name dkm-train-tartanair data_root=./data/datasets trainer.max_epochs=5 datamodule.include="$scene\_$d"
  

Config Overrides

• Example:
  • Use subset: datamodule.include="<regex>" or datamodule.exclude="<regex>"
  • Change split ratio: datamodule.split_ratio=[0.5,0.3,0.2]
  • Change validation interval: trainer.val_check_interval=<interval>