Reproduce “Context-Free TextSpotter for Real-Time and Mobile End-to-End Text Detection and Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ryota Yoshihashi, Tomohiro Tanaka, Kenji Doi, Takumi Fujino, Naoaki Yamashita
Researchers at Yahoo Japan Corporation developed Context-Free TextSpotter, an end-to-end text detection and recognition system designed for real-time, lightweight mobile deployment. The model achieved up to 25 FPS on GPUs and 51ms latency on mobile Neural Engines with only 3.1 million parameters, balancing speed and recognition quality without relying on complex contextual modeling.
Key insights
- It is feasible to achieve competitive E2E text spotting performance on mobile devices with significantly reduced model complexity (3.1M parameters) by simplifying the architecture and avoiding traditional linguistic or geometric contextual modeling.
- A hybrid character spotting approach, combining peak-detection-based (for small characters) and labeling-based (for large characters) methods, enhances the robustness of character localization across varying scales.
- Weakly supervised training, involving an initial approximation of character-level ground truths and a subsequent self-labeling strategy, is an effective method for training text spotters when detailed character annotations are scarce.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.