search

Yuliang-Liu / Curve-Text-Detector

This repository provides train＆test code, dataset, det.&rec. annotation, evaluation script, annotation tool, and ranking.
639 stars 155 forks source link

get pool result? #17

Closed ccnankai closed 6 years ago

ccnankai commented 6 years ago

Thank you for your perfect job. I downloaded the model and put it in the specified folder , run the demo.py ,but the detection location is not as accurate as the image in readme.Position offsets often occur or Missing half of the word on the left and right sides. can you give me some advice?

442 33

Yuliang-Liu commented 6 years ago

Hi there, The results are correct, but I know they look not very well for now. Because for fair comparison and easy reproduction, the method only used provided training data without any tricks, which may restrict its generalization ability to some extent. Using syntext or MLT pre-trained model to finetune on this dataset or testing images with larger scale may be good ways to improve the result.

You can also refer to some recent papers for curve text detection. Our recent solution is a segmentation based method, which can remarkably improve the detection performance and it can be much better in solving "Missing parts" problems.

In a word, the project is a little bit outdated. A new version of curved-text-detector is scheduled to release in the future.

ccnankai commented 6 years ago

@Yuliang-Liu thank you very much