We cannot provide the necessary data for running the model because the OpenEDS dataset has not been released (as of 12th November 2019). However, we provide the code used for winning the OpenEDS Challenge at ICCV 2019.
Please use python 3.5 or above.
- Install requirements
pip install -r requirements.txt
- Download the required files and the dataset
We have no permission to provide the dataset at the moment (12th November 2019)
- Add the directory to PYTHONPATH:
export PYTHONPATH=$PYTHONPATH:$(pwd)
There are two steps for running inference on our model. The first step is already pre-computed and the corresponding files can be downloaded (please see "Setup" above).
-
We identify the most similar images in the unlabeled test set for each test segmentation mask: A segmentation network predicts the semantic eye regions for the unlabeled test set images and we rank these images by their L2 distance to the target segmentation mask. This ranking is already pre-generated and stored in h5 files.
-
For each target segmentation mask in the test set, we then select this "most similar eye image" (one could consider this a nearest neighbour), and refine it via a fully convolutional architecture in order to minimize the final root mean squared error based objective.
In order to run inference and produce a npy file for each test segmentation mask, please run the following command:
python evaluate_refinenet.py
This will produce the submission files in a sub-folder of res/refinenet.