Difference between predictor ran on local and the one ran on hugging face

[ep0p]

I have installed on local the tensorflow version of doctr.
When running the ocr_predictor on local with these parameters ocr_predictor(det_arch='db_resnet50', reco_arch='crnn_vgg16_bn', pretrained=True) i can't seem to get all the text detected:

However, when i am running my document on the hugging face demo with the same models for det and rec, all the text gets detected:

i really need to be able to detect all the text.

[felixdittrich92]

Hi @ep0p thanks for the report :)
Unfortunately the demo isn't up to date with the latest doctr version, but we have already a ticket to fix this

The huggingface demo uses pytorch as backend - could you give it a try with pytorch ?
CC @odulcy-mindee (#1090)

[ep0p]

You were right, pytorch version did it, thank you.

Should i understand that pytorch version is better at detecting and recognizing text or that is not at all the case?

[felixT2K]

With pytorch as backend it works better (all models) yes.
We try always to keep it comparable but sometimes the TF backend makes it hard for us 😅

[ep0p]

Alright, thanks again for your answer

[dlowe24]

Hi, I am trying the same thing on huggingface and Colab and it seems huggingface detect all text while Colab does not capture all text.
I am using the same parameters and the latest doctr v0.9.0 with pytorch. See parameters below.

predictor = ocr_predictor(det_arch="db_resnet50", # Text detection model
reco_arch="crnn_vgg16_bn",assume_straight_pages=True, pretrained=True)

Is this still an issue ?

[felixdittrich92]

Hi @dlowe24 👋,

The HF demo is again out of date 😅 sry for this 🙏
I opened now an automation to avoid this again #1742

[dlowe24]

Thanks. Could indicate which version huggingface was using ?

[felixT2K]

Should be 0.8.0 :)

[dlowe24]

thanks

[felixT2K]

Demo is now up to date with the main branch and will be updated automatically (also after each release) :)

[dlowe24]

Thanks. I can confirm the results are the same as in Colab.
However, what ever model that was used in 0.8.0 seem to be a better model at recognizing all the text in the attached image.