A Dataset and Evaluation Metric for Coherent Text Recognition from Scene Images

In this paper, we deal with extraction of textual information from scene images. So far, the task of Scene Text Recognition (STR) has only been focusing on recognition of isolated words and, for simplicity, it omits words which are too short. Such an approach is not suitable for further processing of the extracted text. We define a new task which aims at extracting coherent blocks of text from scene images with regards to their future use in natural language processing tasks, mainly machine translation. For this task, we enriched the annotation of existing STR benchmarks in English and Czech and propose a string-based evaluation measure that highly correlates with human judgment.