Metrics comparing docTR to other available OCR's

[aditya-ihx]

There is no official metric to show how much we better/worse we perform comparing to other OCR's like pytesseract or Easy-OCR. How about having a metric section for the same. This can be helpful so that others can find docTR more as well as build trust among the users who are skeptical

[felixdittrich92]

Hi @aditya-ihx 👋

We have a metric comparision (with the cloud provider OCR solutions) :)
https://mindee.github.io/doctr/using_doctr/using_models.html#end-to-end-ocr

[SkaarFacee]

Hey, can anyone give me an update about this ?

[felixdittrich92]

On it :)
About the inference latency i have already added some benchmarks in https://github.com/felixdittrich92/OnnxTR

[SkaarFacee]

Oh yes, that looks perfect. Something similar for recall and precision would be perfect 🥳

[gkartik13]

Thanks for the latency comparison! But I was wondering if there are any plans on adding information comparing DocTR to EasyOCR, PaddleOCR and Tesseract in terms of precision and recall?

[felixdittrich92]

Hey @gkartik13 👋,

Currently we have benchmarks against the big cloud provider OCR services (Google, AWS, Azure (Microsoft)): https://mindee.github.io/doctr/using_doctr/using_models.html#end-to-end-ocr

For EasyOCR this would be easy ^^ because it returns the results on word boundaries for PaddleOCR, pytesseract (subline / possible word pair results) / Surya (line boundaries) that's not the case - this would require some additional logic.

So everyone is invited to work on this - we are happy about every contribution - fell free to open an issue for this - but on our end some other topics have a much higher priority atm :)