This branch started out as an investigation into the possibility of adding a second output mask which predicted the bottom of text words, to help with separately closely spaced lines in the text region mask. However I don't think that approach is going to be the best way forwards. Along the way I made some improvements to the training progress on this branch, which I want to land:
Make debug image generation during training and validation optional, controlled by a CLI flag
Add a progress bar during validation. Useful when validating on the full ~1.2K validation images in the HierText dataset
This branch started out as an investigation into the possibility of adding a second output mask which predicted the bottom of text words, to help with separately closely spaced lines in the text region mask. However I don't think that approach is going to be the best way forwards. Along the way I made some improvements to the training progress on this branch, which I want to land: