A suite of tools for robust and generalizable estimation of eye shape from videos.
With pylids you can use a pretrained DNN model based on DLC to:
- Estimate the pupil outline
- Estimate the shape of the eylids
pylids also provides users with tools to finetune the default DNN model to ensure generalization on their dataset. Users can:
- Automatically generate optimally selected domain specific data augmentations to improve pupil and eyelid estimation on new datasets
- Select miniminum frames to relabel from the new dataset to ensure generalization
pylids has been built to be used with the pupil lab gaze estimation pipeline.
Use the shell script pylids_setup.sh
Tested with CUDA version 10.2
Check out the notebooks in the demo folder to see how to use pylids and train new models.
The creation of pylids was funded by NSF EPSCoR # 1920896 to Michelle R. Greene, Mark D. Lescorart, Paul MacNeilage, and Benjamin Balas.