Preparation of training data

GlottisNetV1 and GlottisNetV2

Necessary data for training:

• Images of the glottis

• Segmentation of glottal area

• Annotations of anterior and posterior points

GlottisNetV1 and GlottisNetV2 are trained on the openly available BAGLS data set that can be found in https://github.com/anki-xyz/bagls. The annotations of the anterior and posterior points are provided in https://github.com/anki-xyz/GlottalMidline. The coordinates need to be stored as JSON.

GlottisNetV2 trained on videos

The used videos can be found on Zenodo https://zenodo.org/record/6938457.

The videos are stored as mp4-files and have 30 frames each. For the annotation of the anterior and posterior points, the in <https://github.com/anki-xyz/GlottalMidline>`_ provided tool, was used. The segmentations were created using a trained model of GlottisNetV2 on the frames of the input videos. The corresponding Python script predict_segemntation.py is available in this repository in the directory “Examples”. Those precalculated segmentations were enhanced and revised sing PiPra, provided in https://github.com/anki-xyz/pipra.