This repository is NAVER CLOVA’s well-known deep learning framework for scene text recognition, published alongside the ICCV 2019 paper “What Is Wrong With Scene Text Recognition Model Comparisons?” It has about 3,900 stars on GitHub and remains a standard reference codebase for OCR research and practice.
The project systematically breaks text recognition into four stages (transformation, feature extraction, sequence modeling and prediction) and provides modular implementations of the main choices for each: TPS spatial transformers, ResNet backbones, BiLSTM sequence models and CTC or attention decoders. Because the stages are interchangeable, researchers can run fair comparisons and practitioners can assemble a recognizer suited to their data. The repo includes training and evaluation code, pretrained models for the standard configurations, and support for the common benchmark datasets as well as synthetic training data pipelines such as SynthText and MJSynth. It was last updated in 2024.
The project is fully free and open source under the Apache 2.0 license. There is no commercial edition or hosted version; code and pretrained models are available at no cost.
This benchmark is for OCR researchers and engineers who need a solid, comparable baseline for scene text recognition, or a modular codebase to train recognizers for their own images and documents.