Model Zoo

This page lists the presets and release artifacts documented for version 0.1.12.

Detection

Preset	Architecture	Params	Rotated	Artifacts	Origin	License
`east_50_g1`	EAST (ResNet-50)	53.86M	Yes	east_50_g1.onnx east_50_g1.pth	Manuscript	MIT
`yolo26s_obb_text_g1`	YOLO26-S OBB	9.75M	Yes	yolo26s_obb_text_g1.raw.onnx yolo26s_obb_text_g1.pt yolo26s_obb_text_g1.raw.yaml	Trained by the authors with Ultralytics YOLO	Ultralytics license
`yolo26x_obb_text_g1`	YOLO26-X OBB	57.61M	Yes	yolo26x_obb_text_g1.raw.onnx yolo26x_obb_text_g1.pt yolo26x_obb_text_g1.raw.yaml	Trained by the authors with Ultralytics YOLO	Ultralytics license

Preset	Architecture	Params	Supports	Artifacts	Origin	License
`SimpleSorting`	Algorithmic ordering		Left-to-right, multi-column		Manuscript	MIT

Preset	Architecture	Params	Script	Artifacts	Origin	License
`trba_base_g1`	TRBA	45.10M	Modern + pre-reform Russian	trba_base_g1.onnx trba_base_g1.pth trba_base_g1.json trba_base_g1.txt	Manuscript	MIT
`trba_lite_g1`	TRBA-Lite	9.46M	Modern + pre-reform Russian	trba_lite_g1.onnx trba_lite_g1.pth trba_lite_g1.json trba_lite_g1.txt	Manuscript	MIT
`trba_lite_g2`	TRBA-Lite	9.46M	Modern + pre-reform Russian	trba_lite_g2.onnx trba_lite_g2.pth trba_lite_g2.json trba_lite_g2.txt	Manuscript	MIT

Preset	Architecture	Params	Orthography	Artifacts	Origin	License
`modern_charlm_g1`	CharLM	4.38M	Modern Russian	modern_charlm_g1.onnx modern_charlm_g1.pt modern_charlm_g1.json modern_words.txt	Manuscript	MIT
`prereform_charlm_g1`	CharLM	4.39M	Pre-reform Russian	prereform_charlm_g1.onnx prereform_charlm_g1.pt prereform_charlm_g1.json prereform_words.txt	Manuscript	MIT

EAST: An Efficient and Accurate Scene Text Detector (Zhou et al., CVPR 2017) — академическая основа для семейства детекторов EAST. Реализация в manuscript-ocr основана на оригинальной архитектуре, но процедура обучения существенно переработана. Предобученные веса получены авторами проекта.
What Is Wrong With Scene Text Recognition Model Comparisons? Dataset and Model Analysis — архитектурная основа для семейства TRBA (TPS-ResNet-BiLSTM-Attn). Распознаватели в manuscript-ocr адаптированы под задачи проекта.