FYI on OpenEDS Sparse Semantic Segmentation

OpenEDS Sparse Semantic Segmentation Challenge 20
1
  1. Please find common submission errors and corresponding solutions below.
  • Mismatch of filename!! Check if your filename is correct.
    This happens for cases where the key in submission does not exist. Carefully check your files and re-submit.
  • Invalid data type! Should be uint8!
    We require the data type of predictions to be uint8. Please change the data type and re-submit.
  1. Challenge participation deadline: July 31, 2020
2

Regarding the question of using pre-trained model, below is the general comments:

We welcome using other datasets (not eye) to treat as a warm start of training the model. However, it is NOT allowed to use other eye dataset as the pre-trained model, since we focus on the study of sparse information in sequence data for eye segmentation.

In order to check the validity of the pre-trained models not using eye-dataset for the winners - we would ask for the full code pipeline from scratch training or using the off-the-shelf models.

3

Regarding the question of “usage of unlabeled images”, it would be okay to utilized some tech to generate some pseudo masks for those unlabeled images, as long as they are not manually annotated.

1 Like
4

moved question here from pretrained thread

Are there any constraints to using a technique such as a skew or rotate of the given data to enable more diversity in examples? While this is creating more data, it is unclear if this is considered acceptable.

5

General data augmentation is allowed to boost performance.