Scene Text Recognition (STR) relies critically on large-scale, high-quality training data. While synthetic data provides a cost-effective alternative to manually annotated real data, existing rendering-based synthetic datasets often suffer from a domain gap with real-world text due to insufficient diversity.
UnionST is a strong data engine that synthesizes text covering a union of challenging samples to better align with the complexity observed in the wild. Models trained on the resulting UnionST-S dataset achieve significant improvements over traditional synthetic datasets on challenging STR benchmarks.
The models (such as SVTRv2-AR) are implemented using the
OpenOCR framework. Training can be initiated with:
1cd OpenOCR
2torchrun --nproc_per_node=8 tools/train_rec.py --c configs/rec/nrtr/svtrv2_nrtr_unionst.yml
1@inproceedings{ye2026wrong,
2 title={What's Wrong with Synthetic Data for Scene Text Recognition? A Strong Synthetic Engine with Diverse Simulations and Self-Evolution},
3 author={Ye, Xingsong and Du, Yongkun and Zhang, JiaXin and Li, Chen and LYU, Jing and Chen, Zhineng},
4 booktitle={CVPR},
5 year={2026}
6}
This project is licensed under the MIT License.
