Contextual Text Block Detection Towards Scene Text Understanding

Most existing scene text detectors focus on detecting characters or words that only capture partial messages due to missing contextual information. For a better understanding of in scenes, it is more desired detect blocks (CTBs) which consist one multiple integral units (e.g., characters, words, phrases) natural reading order and transmit certain complete messages. This paper presents detection, new setup detects CTBs for texts scenes. We formulate the by dual detection task first then groups them into CTB. To this end, we design novel clustering technique treats as tokens (belonging same CTB) an ordered token sequence. In addition, create two datasets SCUT-CTW-Context ReCTS-Context facilitate future research, where each CTB well annotated sequence units. Further, introduce three metrics measure local accuracy, continuity, global accuracy. Extensive experiments show our method accurately effectively facilitates downstream tasks such classification translation. The project available at https://sg-vilab.github.io/publication/xue2022contextual/ .