CRAFT: Character-Region Awareness For Text detection

CRAFT ONNX

• Origin repo: https://github.com/clovaai/CRAFT-pytorch

Getting started

Install dependencies

Requirements

• My Environment:
  • pytorch 1.8.1+cu102
  • onnxruntime 1.10.0
  • check requirements.txt

pip install -r requirements.txt

Convert Craftmlt25k.pth to onnx

• Download model pth from origin repo: craftmlt25k and push to folder weights
• Run:

CUDA_VISIBLE_DEVICES=0 python3 craft2onnx.py --craftmlt25kpthpath <craft_model_path.pth> --craftonnxpath <craft_model_onnx_save_path.onnx>

Inference Craftmlt25k without refinet

• Run:

CUDA_VISIBLE_DEVICES=0 python3 infer_craft_without_refinet.py --craftonnxpath <craft_model_onnx_path.onnx> --image <image_path_inference>

Convert Refinet.pth to onnx

• Download model pth from origin repo: refinet and push to folder weights
• Run:

CUDA_VISIBLE_DEVICES=0 python3 craft2onnx.py --craftmlt25kpthpath <craft_model_path.pth> --refinetpthpath <refinet_model_path.pth> --refinetonnxpath <refinet_model_onnx_save_path.onnx>

Inference Craft model with refinet

• Run

CUDA_VISIBLE_DEVICES=0 python3 infer_craft_with_refinet.py --craftonnxpath <craft_model_onnx_path.onnx> --refineonnxpath <refinet_model_onnx_path.onnx> --image <image_path_inference>

Test instruction using pretrained model

• Download the converted models

Model name	Used datasets	Languages	Purpose	Model ONNX Link
General	SynthText, IC13, IC17	Eng + MLT	For general purpose	Click
LinkRefiner	CTW1500	-	Used with the General Model	Click

To do list

• Convert General craft_mlt_25k to ONNX with dynamic shape
• Convert LinkRefiner to ONNX with dynamic shape
• Convert General craft_mlt_25k to TensorRT with dynamic shape
• Convert LinkRefiner to TensorRT with dynamic shape
• Convert IC15 to ONNX with dynamic shape

REFERENCE

1. https://github.com/clovaai/CRAFT-pytorch