👀 Integrate with Amazon Textextract

[awtkns]

Currently the only OCR service tarsier supports is GoogleOCR vision. It would be good to provide another ocr service that allows textextract to be used

[shubhamofbce]

I think we need this asap, because google vision is not working as expected for any complex website. I am working on this.

[awtkns]

@shubhamofbce let me know if you need support!

[plamb-viso]

bump; very interested in testing this library out using textract output

[asim-shrestha]

@plamb-viso happy to take a PR! It should be fairly straightforward as we have this somewhat abstracted.

We'd also really like to test out Azure OCR as we've heard its the most performant. (Will make a separate issue for this)

[asim-shrestha]

And any luck @shubhamofbce ?

[shubhamofbce]

@asim-shrestha Sorry I have not update.
I looked into it long back, it was straight forward but didn't get a chance to complete it and create a PR and now I don't have that with me.

[asim-shrestha]

No worries @shubhamofbce , did you still want to tackle this?

[shubhamofbce]

Sorry, but I will not be able to work on it due to time constraint. @asim-shrestha

[Loeing]

I think I should be able to tackle this next week

[awtkns]

Hey @Loeing let me know if you you need any support on this one.

[Loeing]

@awtkns sorry this past week has been busier than anticipated. Have been playing around with Tarsier. Should be able to make some progress by the end of next week

[tvatter]

@Loeing I'm super interested in the ability to integrate with Amazon Textextract. Have you made any progress on this? Is there any chance I can be of some assistance?

[mscully4]

Howdy! I pulled down the code and tried my hand at integrating with AWS Textract. I ran into a small problem, Textract only returns normalized geometry data (values between 0 and 1), which differs from GCP & Azure. This seems to cause an issue with this line of the format_text method, which checks spacing between annotations using 10 pixels as its baseline. Since the data is normalized, everything gets squished onto one line in the output. De-normalizing the data (multiplying the normalized values by the height/width of the image) fixed the issue and produced correct looking output. The question I have is: would you rather I just de-normalize the Textract response data or should the format_text function be updated to only operation using normalized values?

[philipbjorge]

Does anyone have any published WIP branches available to look at?
Thanks