docTR and onnxTR don't give same results

[lionelfer]

Hi,
I try these 2 demo (Felix92/OnnxTR-OCR and mindee/doctr) with same models (db_resnet50+crnn_vgg16_bn) on this image

I get different results.
docTR perform well, it read: url=http://localhost:8000/ocr/"
onnxTR read: url-http//Localhost:8000/ocr/-
So onnxTR see "-" twice, never read the ", read a "L" instead of "l"

I read the code and didn't point out the pre treatment or the parameters with can explain differences.
I tried a local instance of onnxTR and got the same result as the demo Felix92/OnnxTR-OCR .
I expect to get same results (with same models !) with Felix92/OnnxTR-OCR and mindee/doctr.
Can you help me ? am I missing something ?

[felixdittrich92]

Hi @lionelfer 👋,

Unfortunately the Huggingface doctr demo lacks a bit behind (We have already a ticket in docTR to automate the process that the demo is more up to date mindee/doctr#1090 CC @odulcy-mindee... i tested your image with the given models in docTR and OnnxTR and the results are extactly the same:

Det model: db_resnet50
Mean difference between DOCTR and ONNXTR prob maps: 0.00000115

Render results:

OnnxTR:
url
http://localhost: 8000/ocr/"

docTR: 
url
http://localhost: 8000/ocr/"

Btw.: Passing such small already cropped images in tricky (maybe works better with the recognition_predictor standalone), because the detection models are trained on documents and every input is resized to 1024x1024 before

[felixdittrich92]

Oh ..wrong model (used db_mobilenet_v3_large for the render results) now with db_resnet50 i can reproduce the result in both (docTR & OnnxTR):

url
http - //Localhost: 8000/ocr/ -

[lionelfer]

Thanks @felixdittrich92, your answer is helpful. i will try the recognition_predictor on this kind of image.

[felixdittrich92]

@lionelfer

from onnxtr.io import DocumentFile
from onnxtr.models import recognition_predictor

doc = DocumentFile.from_images("/home/felix/Desktop/test_cropped_2.png")

reco_predictor = recognition_predictor("crnn_vgg16_bn")
res = reco_predictor(doc)

print(res)

[('"http://localhost:8000/ocr/"', 0.5127920508384705)]

Keep in mind we don't recognize whitespaces ! :) Also within the detection on a full image this would be 3 parts: ["url", "=", ""http://localhost:8000/ocr/""]