Benchmarking Automatic Multi-script Scene Component Transcription for AUTNT Dataset

摘要

Followed by tremendous success in text detection from scene imagery and videoframes with the advent of deep learning, transcription of localized text componentshas emerged as an imperative task in end-to-end text reading systems. Though, transcriptionis a part of research in end-to-end system development, focused study ontranscription as a standalone research problem is nonetheless a necessity consideringthe huge amount of associated complexity. In this paper, we present the firstbenchmark performance of multi-script scene as well as document type componentsof a publicly available multi-utility dataset named AUTNT. Besides traditionalmetrics such as word recognition rate or character recognition rate, we include editdistance distribution to reflect more on the obtained results. Moreover, we presentperformances on different subsets and relevant union of them to have more insightinto the obtained state of transcription. Overall recognition rates at component andcharacter level are 72.90% and 86.43% respectively. As the component images ofthis dataset bear with high degree of practical complexity, this result is reasonablyacceptable as the first initial benchmark. Research fraternity is invited to outperformthis benchmark through development of competing methods.