- Improve/Train text segmentation, bubble segmentation models (eg. with color, contrast augmentations since light grey text is usually missed)
- Train own inpainting model
- Fix issue where text seg model connected components fails on large text where characters are very separated
- Fix issue for bubble seg model where same bubble can be predicted for multiple times
- Update Streamlit web app
- Add CLI interface
- Add Docker interface
- Separate text segmentation between sfx and actual text
- Refactor and clean code
- Decide to keep or remove Dectectron2 bubble segmentation model