Performance Evaluation for Text Localization Algorithms: An Empirical Study

Text localization is important for image-based text analysis and has been intensively studied in the last decade. Even though several metrics have been proposed, developing a general performance evaluation metric according to the nature of text localization task is still an open problem. This paper presents an empirical study on several representative performance evaluation metrics by formulating them into a unified framework. To compare these metrics, we report experimental results on the ICDAR 2005 text locating competition datasets with a state-of-the-art text localization algorithm, and based on the performance scores, discuss the different characteristics of these metrics. The performance scores based on different metrics provide a benchmark for evaluating future works.