- first extract layout block in documents
- show UI so the user can edit when the output has mistakes
- for research paper then layout extraction should make few mistakes
- then use the data collect by user to train model
- can we make this task into a game ?