Dataset for Detection + OCR pipeline and comparison with Nomeroff-Net about nomeroff-net HOT 2 CLOSED

At the moment I try to implement my own model for the license plate detection and recognition. I want to compare my model performance with Nomeroff-Net results. Do you have any dataset with non-cropped images annotated with bboxes and license plate numbers so that I can use it in my pipeline (detection + OCR)?

All datasets related to this project are made publicly available. There are no datasets that you are interested in.
For some types of numbers, we bought datasets on the site platesmania, we do not have the right to distribute them, they give xml with links to uncut photos. On these datasets, you can evaluate your and our algorithm.

And also, in your post it is written that you reached ~ 97% word accuracy on Ukrainian numbers. Did you use the test part of this dataset for metrics evaluation? Is 97% accuracy only OCR metrics or you make detection + OCR and then compute metrics? If not, did you try to measure full pipeline (detection + OCR) metrics and what were your results?

The information in this article is already outdated, now the accuracy on Ukrainian OCR is higher (for cropped & normalized numberplates).

We have a small script for taking measurements of what you asked about. There are a small number of numberplates.
https://github.com/ria-com/nomeroff-net/blob/master/examples/ju/benchmark/accuracy-test.ipynb

from nomeroff-net.

@ApelSYN got it! Many thanks!

from nomeroff-net.