4.2.2. CTW1500

It is a challenging text detection dataset in long curve format, 1000 for training and 500 for testing form a total of 1500 images. Unlike traditional text datasets (such as ICDAR2017 MLT), the text instance in CTW1500 is marked by a 14-point polygon. The annotations in this dataset are labelled in textline level, which can describe the arbitrary curved form.

#### *4.3. Performance Evaluation Criteria*

In this detection algorithm, three evaluation indexes are involved, namely:
