A Tri-map free direct Alpha Matting.
An Attention-based Multi-scale Matting Network is a tri-map(side information) free deep image matting network.
Follow the instruction instruction to contact author for the dataset.
I have trained the model using the adobe dataset and provide a pretrained model.
You have to define your customized dataloader based on the files in data_loader directory.
Write dataset class refer to data_loader/create_dataset.py.
Write dataloader class refer to data_loader/data_loaders.py.
$ python train.py -c config.json
$ python train.py --resume /dir/to/the/saveing/checkpoint -c config.json
Finetune from a pretrained checkpoint.
$ python train.py -f ./pretrained.pth -c config.json
If you want to visualize during training, run in your terminal:
$ tensorboard --logdir saved/runs/
Use test.py to test your dataset.
From a best checkpoint of medium-depth network that are trained for 60 epoch.
