FYI on 3D Eye Segmentation Challenge

This is the FYI for the Facebook Eye Tracking 3D Eye Segmentation Challenge

  1. Follow instructions on Facebook at ICCV 2021 - OpenEDS Challenge - Facebook Research to create the required format / prediction data and use the code provided in the Dropbox to create.json file in order to submit to EvalAI for model evaluation.

Example of the folder to save predictions:
------ (main-folder) predictions
|--------- output.txt
|--------- (sub-folder) PID_0
|-------------------- pred.npy
|--------- (sub-folder) PID_1
|-------------------- pred.npy
|--------- (sub-folder) PID_2
|-------------------- pred.npy

  1. The evaluation metric on the challenge is the mean intersection-over-union for: pupil, iris, sclera, lashes. Background is not included in the evaluation.

  2. Common submission errors and corresponding solutions:

  • assert pred.shape == gt.shape
    This happens for cases where the dimension of predictions and of GT (mask). Please go back and check your results.
  • assert gt== pred
    This happens for cases where the filename of predictions is not the same as GT filename. Please revisit the challenge website and make sure the generated data filenames match our requirement.