FGI-Matting

The official repository for Deep Image Matting with Flexible Guidance Input.

Paper: https://arxiv.org/abs/2110.10898

Requirements

• easydict
• numpy
• opencv-python
• Pillow
• PyQt5
• scikit-image
• scipy
• toml
• torch>=1.5.0
• torchvision

Models and supplementary data for DIM test set(Composition-1k) and Distinctions-646 test set

Google drive: https://drive.google.com/drive/folders/13qnlXUSKS5HfkfvzdMKAv7FvJ6YV_wPK?usp=sharing
百度网盘： https://pan.baidu.com/s/1ZYcbwyCIrL6G9t7pkCIBYw 提取码: zjtj

• Weight_DIM.pth The model trained with Adobe matting dataset.

• Weight_D646.pth The model trained with Distincions-646 dataset.

• DIM_test_supp_data.zip Scribblemaps and Clickmaps for DIM test set.

• D-646_test_supp_data.zip Scribblemaps and Clickmaps for Distinctions-646 test set.

Place Weight_DIM.pth and Weight_D646.pth in ./checkpoints.
Edit ./config/FGI_config to modify the path of the testset and choose the checkpoint name.

Test on DIM test set(Composition-1k)

Methods	SAD	MSE	Grad	Conn
Trimap test	30.19	0.0061	13.07	26.66
Scribblemap test	32.86	0.0090	14.18	29.09
Clickmap test	34.67	0.0112	15.45	30.96
No guidance test	36.36	0.0141	15.23	32.76

"checkpoint" in ./config/FGI_config.toml should be "Weight_DIM".
bash test.sh
Modify "guidancemap_phase" in ./config/FGI_config.toml to test on trimap, scribblemap, clickmap and No_guidance.
For further test, please use the code in ./DIM_evaluation_code and the predicted alpha mattes in ./alpha_pred.

Test on Distinctions-646 test set(Not appear in the paper)

Methods	SAD	MSE	Grad	Conn
Trimap test	28.90	0.0105	24.67	27.40
Scribblemap test	33.22	0.0131	26.93	31.38
Clickmap test	34.97	0.0146	27.60	33.11
No guidance test	36.83	0.0156	28.28	34.90

"checkpoint" in ./config/FGI_config.toml should be "Weight_D646".
bash test.sh
Modify "guidancemap_phase" in ./config/FGI_config.toml to test on trimap, scribblemap, clickmap and No_guidance.
For further test, please use the code in ./DIM_evaluation_code and the predicted alpha mattes in ./alpha_pred.

The QT Demo

Copy one of the pth file and rename it "Weight_qt_in_use.pth", also place it in ./checkpoints.
Run test_one_img_qt.py. Try images in ./testimg. It will use GPU if avaliable, otherwise it will use CPU.

I recommend to use the one trained on DIM dataset.
Have fun :D

Acknowledgment

GCA-Matting: https://github.com/Yaoyi-Li/GCA-Matting