jeromepatel/Eye-Segmentation-3rd-Rank-Winning-Solution

It contains reproducible code for 3rd winner of OpenEDS 2021 Track 1 contest - The method is based on PointNet 2

3D Eye Segmentation - 3rd Rank Winning Solution

Instructions:

There are two separate scripts for Training and Testing.

Training:

• Training can be done using train_eyesemseg.py. To use CPU when training, use --gpu 0and by default GPU option is enabled.
• Place the dataset in data/eyeseg directory, please make sure it includes train and val part inside them.
• Install all the necessary libraries mentioned inside requirements.txt.
• Simply run the command python train_eyesegmseg.py for default arguments, which include gpu 1 and epoch 32 too.

The model will train for 32 epochs and will save the pretrained model inside log/eyeseg/* directory.

Inference:

• Testing can be done using test_eyesemseg.py. GPU and CPU option are available for testing as well, using gpu 0.
• Important arguments are --log_dir pre_trained_model/2021-08-14_19-08, --test_dir test/val --save_labels True/False. --valid can be used if the labels for the test_dir are present inside the directory(eg. in case of val)
• Finally to run the inference from pre_trained_model (provided by us), run the command python test_eyesemseg.py --log_dir pre_trained_model --test_dir test --save_labels True.
• The labels can be found as *.npy files inside output folder.
• To create json file from npy labels, use create_json_ss.py script by running the command python create_json_ss.py --list-file submissionFiles.txt --submission-json pointnet2Submission.json

Credits: This repo contains modified version of Model and Training script from this repo