The project refers to https://github.com/lizhengwei1992/Semantic_Human_Matting.
Finally, I converted the pytorch model to the openvino model, and run it on cpu.
python == 3.6.5
pytorch-gpu == 1.2.0
torchvision == 0.4.0
cv2 == 4.1.2
openvino == 2019.3.334
opencv == 4.1.2
Download alpha images from: https://pan.baidu.com/s/1geFwrv7H3moFc9nxXwQJ1g (boy2)
Download init images and mask images from: https://pan.baidu.com/s/1R9PJJRT-KjSxh-2-3wCGxQ (dzsn)
Unzip alpha to data folder and execute:
(1) cd data
(2) use alpha images to generate trimap images: python gen_trimap.py
(3) generate train.txt: python gen_train_data_list.py
I set patch_size = 256, train_batch = 32.
(1) train phase 1:
python train.py --finetuning False --lr 1e-3 --nEpochs 100 --train_phase pre_train_t_net
(2) train phase 2:
python train.py --finetuning True --lr 1e-4 --nEpochs 200 --train_phase end_to_end
Put images into images folder and run python test_image.py.
(1) Convert model_obj.pth to model_obj.onnx: python torch_to_onnx.py
(2) Copy model_obj.onnx to C:\IntelSWTools\openvino_2019.3.334\deployment_tools\model_optimizer, and run:
python mo_onnx.py --input_shape=[1,3,256,256] --mean_values [104.0,112.0,121.0] --scale_values [255.0,255.0,255.0] --input_model .\model_obj.onnx -o .\
Alter that, we have two files: model_obj.xml(model information file) and model_obj.bin(weights file).
Build Kit: Desktop Qt 5.12.0 MSVC2015 64bit, Release.
OpenVINO model files: openvino_model/model_obj.xml, openvino_model/model_obj.bin.
(1) Create a new qt console project.
(2) Copy the content of file 'for_qt/console_test.pro' to your own .pro file.
(3) Use file 'for_qt/main.cpp' in your project.