mohamadzeina / moby_eye_tracking Goto Github PK

Lots of variables are passed between them inelegantly, would be much simpler if they were part of one object

When loading a batch of images to train on, the script should check whether they were taken with a central webcam. This can either be extracted from a .txt file uploaded in the same folder that describes the configuration, or there could be a specific suffix in the directory name for the batch of images.

This should significantly improve accuracy, as some techniques rely purely on this pupil localisation with little or no calibration.

I am currently training on the entire dataset, and not performing any validation. There should be a separate validation dataset, either of me at a different angle/with different lighting, or of a different person.

Including:

• gaze preview points per training point
• folder name to save images into

For assessing a specific model, without adding data