TextDC: Exploring Multidimensional Text Detection via a New Benchmark and Solution

Text detection has been significantly boosted by the development of deep neural networks but most existing methods focus on a single kind text instance (i.e., overlaid text, layered scene text). In this paper, we expand task from dimension to multiple dimensions, thus providing multi-type descriptions for and content analysis videos. Specifically, establish new detect classify instances simultaneously, termed TextDC. As far as know, benchmarks cannot meet requirements proposed task. To end, collect large-scale classification dataset, named Text3C, which is annotated using multilingual labels, location information, categories. Together with collected introduce multi-stage strict evaluation metric, penalizes approaches missing instances, false positive detection, inaccurate boxes, error categories, developing benchmark TextDC addition, extend several state-of-the-art detectors modifying prediction head solve Then, generalized framework designed formulated. Extensive experiments updated are conducted established verify solvability task, challenges effectiveness solution.