Show, Attend and Read: A Simple and Strong Baseline for Irregular Text Recognition

Recognizing irregular text in natural scene images is challenging due to thelarge variance in text appearance, such as curvature, orientation anddistortion. Most existing approaches rely heavily on sophisticated modeldesigns and/or extra fine-grained annotations, which, to some extent, increasethe difficulty in algorithm implementation and data collection. In this work,we propose an easy-to-implement strong baseline for irregular scene textrecognition, using off-the-shelf neural network components and only word-levelannotations. It is composed of a $31$-layer ResNet, an LSTM-basedencoder-decoder framework and a 2-dimensional attention module. Despite itssimplicity, the proposed method is robust and achieves state-of-the-artperformance on both regular and irregular scene text recognition benchmarks.Code is available at: https://tinyurl.com/ShowAttendRead