CAT-Net: Learning Canonical Appearance Transformations

Code to accompany our paper "How to Train a CAT: Learning Canonical Appearance Transformations for Direct Visual Localization Under Illumination Change".

Dependencies

• numpy

• pytorch + torchvision (0.3.0)

• PIL

• visdom

• pyslam + liegroups (optional, for running odometry/localization experiments)

Running the demo experiment

1. Download the ETHL dataset from here.

2. Update run_cat_experiment.py to point to the appropriate local paths.

3. In a terminal run python3 -m visdom.server -port 8097 to start the visualization server.

4. In another terminal run python3 run_cat_experiment.py to start training.

5. Tune in to localhost:8097 and watch the fun.

Running the localization experiments

Note: the scripts referenced here are from an older version of the repository and may need some adjustments. 1. Ensure the pyslam and [liegroups] packages(https://github.com/utiasSTARS/liegroups) are installed 2. In a terminal open the localization directory and run python3 run_localization_[dataset].py 3. You can compute localization errors against ground truth using the compute_localization_errors.py script.

Pre-trained models

Coming soon!

Citation

If you use this code in your research, please cite:

@article{2018_Clement_Learning,
  author = {Lee Clement and Jonathan Kelly},
  journal = {{IEEE} Robotics and Automation Letters},
  link = {https://arxiv.org/abs/1709.03009},
  title = {How to Train a {CAT}: Learning Canonical Appearance Transformations for Direct Visual Localization Under Illumination Change},
  year = {2018}
}