model: test OCR model frameworks

[simra]

todo: add notes from hackathon

See also this post:

https://www.linkedin.com/posts/farid-hassainia-ca_interested-in-ocr-here-is-an-excellent-library-activity-6884922531560620032-onSr

[monkeypants]

if we have a data architecture that says:

• an image contains zero or more headstamp boundaries
• an agent (person | algorithm) detects zero or more identifying marks (in a given bounding shape, with a given orientation) in a headstamp boundary
• an agent has a qualification

Then we can have many agents making observations in the same images, and we can organise them based on the qualifications of the agent. This agent is a member of the public with (qualification) "user_karma=0.123", that agent is a algorithm with (qualification) "RnD_ignoreme=true", and that other agent has qualifications "expert_group=true" and "expert_group_bulgaria=true"... or something like that.

my point is, we should be able to accommodate multiple OCR model (and human) agents at the same time.

[simra]

Note we've added an mmocr repo where we can tool around with training pipelines. One of our hackers this week did a bunch of work to enable fine-tuning and evaluation of textSnake, which can extract curved text instances.