State-of-the-Art in Action: Unconstrained Text Detection

In this paper, we stage five real-world scenarios for six state-of-the-art text detection methods in order to evaluate how competent they are with new data without any training process. Moreover, this paper analyzes the architecture design of those methods to reveal the influence of pipeline choices on the detection quality. The setup of experimental studies are straight-forward: we collect and manually annotate test data, we reimplement the pretrained models of the state-of-the-art methods, then we evaluate and analyze how well each method achieve in each of our collected datasets. We found that most of the state-of-the-art methods are competent at detecting textual information in unseen data, however, some are more readily used for real-world applications. Surprisingly, we also found that the choice of a post-processing algorithm correlates strongly with the performance of the corresponding method. We expect this paper would serve as a reference for researchers as well as application developers in the field. All collected data with ground truth annotation and their detected results is publicly available at our Github repository: https://github.com/chupibk/HBlab-rlq19.