IndicSTSpotter — model weights

Detector + recognizer for IndicSTSpotter, a unified multilingual scene-text spotter covering 11 Indic scripts + Latin with a single detector, a single recognizer, and one frozen 1,638-character dictionary — no script identification, no per-language weights.

Private while the paper is under review.

Contents

Path	Model	Params	Notes
rec_stage2/	recognizer (SVTR + CTC head, PP-OCRv6_medium_rec)	15.8M	synthetic pretrain → real fine-tune; the paper's headline recognizer
det_mist/	detector (DB, PP-OCRv6_medium_det)	15.5M	trained on real + MIST; used for dense-scene detection & end-to-end
unified_dict.txt	—	—	the frozen 1,638-char dictionary shared by all recognizers

Usage

from huggingface_hub import snapshot_download
d = snapshot_download("MahimaSinghal/IndicSTSpotter", repo_type="model")

from paddleocr import TextRecognition, TextDetection
rec = TextRecognition(model_name="PP-OCRv6_medium_rec", model_dir=f"{d}/rec_stage2")
det = TextDetection(model_name="PP-OCRv6_medium_det", model_dir=f"{d}/det_mist")

For dense scenes, run the detector at a higher inference resolution (long side 1920). See the code repository's REPRODUCE.md for the exact per-table evaluation commands.

License

MIT (code). Model weights released for research use.