aster.pytorch

Why we included this project

This PyTorch port of the ASTER scene text recognizer is worth a look if you work on OCR for photos where text is curved or shot at an angle. Its main selling point is the flexible rectification stage, which straightens distorted text before recognition, and that is the part of scene-text OCR that plain crop-and-classify pipelines tend to struggle with. The repo includes training and test scripts, a pretrained model, and reproduced accuracy numbers on standard benchmarks such as IIIT5k and SVT, so you can check the behavior before adapting it. One caveat: the bidirectional attention decoder from the original paper is not included, and the author is upfront that results may not reproduce under modified settings. Treat it as a research-grade baseline to build on, not a drop-in production service.