Reproduce “SwinTextSpotter: Scene Text Spotting via Better Synergy between Text Detection and Text Recognition” on OpenResearch

About this paper

Mingxin Huang, Yuliang Liu, Zhenghao Peng, Chongyu Liu, Dahua Lin, Shenggao Zhu, Nicholas Yuan, Kai Ding and 1 more

An end-to-end scene text spotting framework, SwinTextSpotter, employs a Recognition Conversion (RC) module and a Dilated Swin-Transformer to enhance the explicit synergy between text detection and recognition. The system achieves state-of-the-art performance across multiple benchmarks for arbitrarily-shaped, multi-oriented, and multi-lingual texts, while eliminating the need for complex rectification modules or character-level annotations.

Key insights

• Explicitly coupling detection and recognition through differentiable loss propagation (like the RC module) fundamentally enhances the synergy between tasks, leading to more robust overall performance.
• Transformer-based architectures and query-based set-prediction approaches can effectively model complex text instances and relationships in end-to-end scene text spotting.
• Achieving state-of-the-art text spotting performance is possible with a simplified architecture, eliminating the need for specialized rectification modules or character-level annotations.

Summary by alphaXiv.