Simple-TMF-Matting

Inofficial simplified version of TMF-Matting with minimal dependencies.

Given an image and a trimap, compute its alpha matte.

Input image	Input trimap	Output alpha matte

The test image is from https://alphamatting.com/datasets.php.

Usage

1. Install PyTorch
2. pip install pillow
3. git clone https://github.com/99991/Simple-TMF-Matting.git
4. cd Simple-TMF-Matting
5. Download the pretrained comp1k.pth model from the original authors' repository and place it in this directory.
6. python test_single_image.py # downloads test images and computes alpha matte

Citing

If you find TMFNet useful in your research, please consider citing the original authors:

@article{jiang2023trimap,
  title={Trimap-guided feature mining and fusion network for natural image matting},
  author={Jiang, Weihao and Yu, Dongdong and Xie, Zhaozhi and Li, Yaoyi and Yuan, Zehuan and Lu, Hongtao},
  journal={Computer Vision and Image Understanding},
  volume={230},
  pages={103645},
  year={2023},
  publisher={Elsevier}
}

Testing on Composition-1K Test Set

1. Download the pretrained model as above.
2. Ask Brain Price to send you Adobe_Deep_Matting_Dataset.zip and place it in this directory. Do not unzip.
3. Download and extract the images of the Pascal VOC2012 dataset to the directory PascalVOC2012. You can also link them with ln -s YOUR_PASCAL_DIR/VOCdevkit/VOC2012/JPEGImages/ PascalVOC2012 if you already have them somewhere else.
4. Run test_composition_1k_dataset.py

Results

MSE × 1000	SAD / 1000
4.547	22.410

MSE is slightly worse and SAD is slightly better than original, but minor details such as background interpolation method result in a large difference, so this is probably acceptable.