search

hwalsuklee / awesome-deep-text-detection-recognition

A curated list of resources for text detection/recognition (optical character recognition ) with deep learning methods.
Apache License 2.0
2.51k stars 513 forks source link

ASTER is not end-to-end in the normal sense #5

Open jdhao opened 5 years ago

jdhao commented 5 years ago

End-to-end recognition means that the whole image is feed into the network and the network will output the recognition result for the whole image.

The input image to the network in ASTER is not the whole image, but a small part containing the warped text. I think it is more proper to call ASTER a recognition algorithm which can deal with irregular text images.

hwalsuklee commented 5 years ago

Thank you for pointing out.

As you mentioned, ASTER is a recognition algorithm. So you can see ASTER in 'Text Recognition' part.

However, A paper of ASTER also showed end-2-end performance for 2 stage approach, ASTER (recognition) followed by TextBoxes (detector).

That's why I added ASTER in 'End-to-End Text Recognition' part also. Please see TABLE 8 in ASTER paper. (http://122.205.5.5:8071/UpLoadFiles/Papers/ASTER_PAMI18.pdf)

jdhao commented 5 years ago

But Table 8 can not justify that ASTER is an end-to-end method. It is coupled with TextBoxes. ASTER is just used for recognizing the text boxes.

End-to-end method means that the detection and recognition are performed by a single method, not cascading of two different methods from two different papers.