[datasets/benchmark] Improve research use

[felixdittrich92]

🚀 The feature

This request can be split into three parts:

• ensure that any already integrated dataset which has the information (boxes & text labels) to be used for the recognition task can also be used for this (crop boxes with corresponding labels)
• add a section in the documentation (like models) for datasets (also split into detection / recognition)
• integrate a script for benchmarking into references/ or directly into the training script for detection and recognition which follows current papers / common used benchmarking splits

detection:
TODO
train:COCO-Text/ ... ?
val: IC03/IC13/...?

recognition:
train: MJSynth/SynthText
val: SVHN/SVT/IIIT5K/IC03/IC13 (+Funsd/Cord)

Motivation, pitch

It would be great to get a comparison to other implementations or other OCR applications for research purposes.
This would make the entire library or its implemented models a little more transparent and easier to compare with others.
As a final point, I have to add that it's just great to see if an implementation reach better benchmarks as other 😅

Additional context

Any feedback or suggestion is very welcome 💯

[felixdittrich92]

closed with #933