In this repo, we use shared matting to replace knn matting method, caused by the error of knn matting. Furthermore, the knn matting is too slow.
numpy==1.14
tensorflow==1.12.0
opencv-python
scipy
imageio
scikit-learn1.1. Download the images and mattes from alphamatting.com.
1.2. Download the MSCOCO dataset, then uncompress to any folder.
-
Modify the
config.pyto adapt to your customed environment and parameters. -
Run
python gen_alpha_from_closed_form.pyto generate the closed form mattes. -
Run
python gen_alpha_from_sharedmatting.pyto generate the shared matting mattes( Replace the knn matting). -
Run
python data_prepare.py --stage=gen_composite_imagesto generate the composite images. -
RUN
python data_prepare.py --stage=gen_image_patchesto generate the image patches. -
RUN
python data_prepare.py --stage=train_val_test_splitto split the train, val, test datasets.
Then we use these patches to train this model. There are totally 270000 patches for all images.
- RUN
python train.pyto train this model, the log and ckpt will be stored in ./logs and ./ckpt
- RUN
python demo.pyto test one image matting.
