With mixed results! I didn’t notice that this page was upside down until after I had uploaded it. The transcription was pretty good (especially sinc the letter was also lopped off on one edge), but Leo got very cinfused at one point and started repeating himself – not the whole letter, just the last half.
Here’s the link:
https://www.tryleo.ai/document/37cbcc2a-b8f9-495b-b6c8-fd530432c298
And here’s a screen shot of part of the repeating bits.
Thanks Tim! When we have access to more computing resources we’re going to mix in examples of rotated/ otherwise distorted images into the training data. It’ll be more expensive but it should straightforwardly resolve problems like these.
Great – I have certainly seen many letters (though not in the batch I’m working with this month) where people write every which way on the same page – so improving Leo’s skill set in that regard might be worht the cost. Not to mention cross-writing! (over the top of what’s already been written, at a 90 degree angle!), to save on postage. If Leo can manage that you should win an award!
Thanks for this. I’m having similar issues. I have multi-page PDF scans that I upload (each individual PDF a document). The PDFs look fine on my computer–all pages are correctly oriented. But after I upload the file, one or more pages ends up turned sideways. If even a single page is turned, LEO’s AI only seems to read first page (or pages) before the turned page and produce a transcript that seems to have content out of order (and does not include the turned page or any pages thereafter). I think there are two issues here: 1-why do multi-page PDFs that look fine before uploading end up with some of the pages out of oriengation? and 2-Is there a workaround, either with LEO AI figuring out how to read pages that aren’t oriented correctly, or to ensure that document uploads maintain orientation integrity?
I also had a similar issue: one of my images, from a letter in French, was upside down, and Leo began translating it as an oddly poetic, if not entirely logical, English-language document before falling into a doom loop.
Interestingly, another letter, this one photographed sideways, got the first line perfectly correct, but then fell into a doom loop:
Hi Kirt, if you could email me with an attachment of those PDFs, it’d be really helpful. We’ll pass them on to the developer who’ll hopefully be able to see what’s going on and refine the mechanism for extracting images from PDFs. Our plan is to introduce image manipulation features soon, so that orientation can be controlled from within the web-app.
This one is very odd!
Image manipulation within the app will be great!
I’ve had similar experiences with upside-down text. Leo transcription often/usually falls into the “doom loop” others are reporting when I’ve neglected to correct the orientation before uploading. Agree very much with Timothy Alborn that image manipulation within the app would be a valuable upgrade.
