This repo reimplements DCNN for image matting.

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.

Usage

numpy==1.14
tensorflow==1.12.0
opencv-python
scipy
imageio
scikit-learn

1.1. Download the images and mattes from alphamatting.com.

1.2. Download the MSCOCO dataset, then uncompress to any folder.

Modify the config.py to adapt to your customed environment and parameters.
Run python gen_alpha_from_closed_form.py to generate the closed form mattes.
Run python gen_alpha_from_sharedmatting.py to generate the shared matting mattes( Replace the knn matting).
Run python data_prepare.py --stage=gen_composite_images to generate the composite images.
RUN python data_prepare.py --stage=gen_image_patches to generate the image patches.
RUN python data_prepare.py --stage=train_val_test_split to 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.py to train this model, the log and ckpt will be stored in ./logs and ./ckpt