search

firebase / quickstart-ios

Firebase Quickstart Samples for iOS
https://firebase.google.com
Apache License 2.0
2.8k stars 1.47k forks source link

Different results when running both iOS and Android versions of MLKit Text Recognition On-Device API with the same image as input #511

Closed joaomelo81 closed 6 years ago

joaomelo81 commented 6 years ago

When using the same image as input for the iOS and Android versions of MLKit Text Recognition On-Device API, I do get different results on the bounding boxes information (x,y,width,height) provided by each solution.

Please see below the original input image I've used for my tests and another image that shows the resulting text block's bounding boxes that were drawn based on the information provided by both of the Text Recognition on-device APIs (in blue is the iOS result and in red is the Android one):

Original input image

Resulting bounding boxes

Does anyone knows what could cause such differences between the results for iOS and Android versions of the API? I suspect they use different ML models for the text recognition / extraction of bounding boxes. If so, is there any chance of having both solutions running the same model in a near future since they are still on beta release?

Any thoughts are welcome!

morganchen12 commented 6 years ago

Filed b/112423460 with the ML team, I'll let you know what they say. I suspect this is a bug or artifact of some offline model packaging caveats.

morganchen12 commented 6 years ago

This issue is the result of the iOS and Android SDKs using different versions of the same OCR engine. It will be resolved in a future update, though not soon. In the meantime, both SDKs should continue to identify text correctly, though they may report slightly different bounding boxes.

If either SDK does not recognize text correctly, please open a new issue and I'll take a look.

Thanks!