This repository contains a PyTorch implementation of the BMVC 2020 paper "Learning To Pay Attention To Mistakes", 2020.

Mou-Cheng Xu is the developer of the code.

1. Download the whole repo including both the code and the datasets folder, compile your environment using bmvc2020_environment.yml file.
2. Use Run.py to run and debug.

How to use this repo with your own datasets:

1. An example of the folder structure of training data is in datasets folder.
2. After you prepare the datasets, you can easily tune the interface in Run.py.
3. In ''network'' argument in ''Run.py'', we provide different combinations of our model to be called: ''ERF_encoder_fp'', ''ERF_encoder_fn'', ''ERF_decoder_fp'', ''ERF_decoder_fn'', ''ERF_all_fp'', ''ERF_all_fn''.
4. When you use any configurations including 'fn' in ''network'', please set the ''reverse'' flag as ''True''.
5. We recomment these two configurations: ''ERF_all_fn'' and ''ERF_all_fp''.

More difficult testing:

To test the generalisation of the trained model on unseen data, we add adversarial noises when we evaluate IoU and Hausdorff distance. This is to compensate the simplicity of the experimental settings (aka. binary segmentation).

Future work:

1. Extending the framework to multi-class.
2. The framework has a great potential in other tasks which need heavy regularisation, such as adversarial attack, denoising, deblurring and semi-supervised learning.

Citation:

If you find the code or the README useful for your research, please cite our paper:

@article{LearnMistakeAttention,
  title={Learning To Pay Attention To Mistakes},
  author={Xu, Mou-Cheng and P. Neil, Oxtoby and C. Alexander, Daniel and Jacob, Joseph},
  journal={BMVC},
  year={2020},
}