Contextual Text Detection

Most existing scene text detectors focus on the detection of characters or words which capture partial textual messages only in most cases due to the missing of contextual information. For a better understanding of text in scenes, it is more desired to detect contextual text blocks which consist of one or multiple integral text units (e.g., characters, words, or phrases) in a specific order, delivering certain independent and complete textual messages. This paper presents Contextual Text Detection, a new setup that detects contextual text blocks for better understanding of texts in scenes. We formulate the new setup by a dual detection task that first detects integral text units and then groups them into a contextual text block. Specifically, we design a novel scene text grouping technique which treats each integral text unit as a token and groups multiple integral tokens belonging to the same contextual text block into an ordered token sequence. To facilitate the future research, we create two new datasets SCUT-CTW-Context and ReCTS-Context where each contextual text block is well annotated by an ordered sequence of integral text units. In addition, we introduce three evaluation metrics that measure contextual text detection in local accuracy, continuity, and global accuracy, respectively. Extensive experiments show that the proposed method detects contextual text blocks effectively. This development including codes, datasets and annotation tools will be published at http://xxxxxxx.