Out-of-Vocabulary Challenge Report

This paper presents final results of the Out-Of-Vocabulary 2022 (OOV) challenge. The OOV contest introduces an important aspect that is not commonly studied by Optical Character Recognition (OCR) models, namely, recognition unseen scene text instances at training time. competition compiles a collection public datasets comprising 326,385 images with 4,864,405 instances, thus covering wide range data distributions. A new and independent validation test set formed are out vocabulary was structured in two tasks, end-to-end cropped respectively. thorough analysis from baselines different participants presented. Interestingly, current state-of-the-art models show significant performance gap under newly setting. We conclude dataset proposed this challenge will be essential area to explored order develop achieve more robust generalized predictions.