GCA-Matting: Natural Image Matting via Guided Contextual Attention

The source codes and models of Natural Image Matting via Guided Contextual Attention which will appear in AAAI-20.

Matting results on test data from alphamatting.com with trimap-user.

Requirements

Packages:

• torch >= 1.1
• tensorboardX
• numpy
• opencv-python
• toml
• easydict
• pprint

GPU memory >= 8GB for inference on Adobe Composition-1K testing set

Models

The models pretrained on Adobe Image Matting Dataset are covered by Adobe Deep Image Mattng Dataset License Agreement and can only be used and distributed for noncommercial purposes.

• ResNet34_En_nomixup: Model of the customized ResNet-34 backbone trained on ImageNet. Save to ./pretrain/. The training codes of ResNet34_En_nomixup and more variants will be released as an independent repository later. You need this checkpoint only if you want to train your own matting model.
• gca-dist: Model of the GCA Matting in Table 2 in the paper. Save to ./checkpoints/gca-dist/.
• gca-dist-all-data: Model of the GCA Matting trained on both Adobe Image Matting Dataset and the Composition-1K testing set for alphamatting.com online benchmark. Save to ./checkpoints/gca-dist-all-data/.

(We removed optimizer state_dict from gca-dist.pth and gca-dist-all-data.pth to save space. So you cannot resume the training from these two models.)

Run a Demo on alphamatting.com Testing Set

python demo.py \
--config=config/gca-dist-all-data.toml \
--checkpoint=checkpoints/gca-dist-all-data/gca-dist-all-data.pth \
--image-dir=demo/input_lowres \
--trimap-dir=demo/trimap_lowres/Trimap3 \
--output=demo/pred/Trimap3/

This will load the configuration from config and save predictions in output/config_checkpoint/*. You can reproduce our alphamatting.com submission by this command.

Train and Evaluate on Adobe Image Matting Dataset

Data Preparation

Since each ground truth alpha image in Composition-1K is shared by 20 merged images, we first copy and rename these alpha images to have the same name as their trimaps. If your ground truth images are in ./Combined_Dataset/Test_set/Adobe-licensed images/alpha, run following command:

./copy_testing_alpha.sh Combined_Dataset/Test_set/Adobe-licensed\ images

New alpha images will be generated in Combined_Dataset/Test_set/Adobe-licensed images/alpha_copy

Configuration

TOML files are used as configurations in ./config/. You can find the definition and options in ./utils/config.py.

Training

Default training requires 4 GPUs with 11GB memory, and the batch size is 10 for each GPU. First, you need to set your training and validation data path in configuration and dataloader will merge training images on-the-fly:

[data]
train_fg = ""
train_alpha = ""
train_bg = ""
test_merged = ""
test_alpha = ""
test_trimap = ""

You can train the model by

./train.sh

or

OMP_NUM_THREADS=2 python -m torch.distributed.launch \
--nproc_per_node=4 main.py \
--config=config/gca-dist.toml

For single GPU training, set dist=false in your *.toml and run

python main.py --config=config/*.toml

Evaluation

To evaluate our model or your own model on Composition-1K, set the path of Composition-1K testing and model name in the configuration file *.toml:

[test]
merged = "./data/test/merged"
alpha = "./data/test/alpha_copy"
trimap = "./data/test/trimap"
# this will load ./checkpoint/*/gca-dist.pth
checkpoint = "gca-dist" 

and run the command:

./test.sh

or

python main.py \
--config=config/gca-dist.toml \
--phase=test

The predictions will be save to** ./prediction by default, and you can evaluate the results by the MATLAB file ./DIM_evaluation_code/evaluate.m in which the evaluate functions are provided by Deep Image Matting. Please do not report the quantitative results calculated by our python code like ./utils/evaluate.py or this test.sh in your paper or project. The Grad and Conn functions of our reimplementation are not exactly the same as MATLAB version.

Citation

If you find this work or code useful for your research, please cite:

@inproceedings{li2020natural,
  title={Natural image matting via guided contextual attention},
  author={Li, Yaoyi and Lu, Hongtao},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={34},
  pages={11450--11457},
  year={2020}
}