Evaluation of Methods for Detection and Localization of Text in Video

The detection and recognition of text from unconstrained, general-purpose video is an important research problem with multiple applications in the surveillance, archiving and content-based retrieval contexts. Many text detection and localization algorithms have been proposed in the literature. However many of these algorithms either make simplistic assumptions as to the nature of the text to be found, or restrict themselves to a subclass of the wide variety of text that is observed in general purpose video. Almost all algorithms operate on images or individual video frames. It is also observed that the published results of most of these algorithms consist simply of sample images with bounded text boxes. There is a need for a quantitative evaluation of these algorithms against a challenging dataset. In this paper we present an evaluation of select text detection and localization algorithms. We present an evaluation of five algorithms. Some of these have been modified from the original work published by the authors. We discuss the method adopted for the evaluation and present results for the text localization methods. We observe that no one text detection and localization method is robust for detecting all kinds of text. It may be necessary to apply different methods that use independent heuristics to extract different kinds of text and then fuse these results temporally and across various algorithms.