clovaai/deep-text-recognition-benchmark

Implements a modular four-stage STR (Scene Text Recognition) framework—Transformation, Feature Extraction, Sequence Modeling, and Prediction—enabling controlled benchmarking of component contributions across accuracy, speed, and memory. Built on PyTorch with support for diverse architectures (TPS/ResNet/BiLSTM/Attention), CTC loss, and standardized LMDB datasets across seven evaluation benchmarks (IIIT5K, SVT, IC03-15, SVTP, CUTE80). Provides pretrained models and comprehensive ablation analysis to isolate performance gains from individual architectural modules rather than conflating improvements across multiple simultaneous changes.

3,926 stars. No commits in the last 6 months.