中文说明

An unofficial implementation of Non-Local Deep Features for Salient Object Detection.

The official Tensorflow version: NLDF

Some thing difference:

1. dataset
2. score with one channel, rather than two channels
3. Dice IOU: boundary version and area version

• Python 3
• Pytorch 1.0
• torchvision
• visdom (optional for visualization)

The information of Loss:

Performance:

Note:

1. only training 200 epoch, larger epoch may nearly the original paper
2. This reproduction use area IOU, and original paper use boundary IOU
3. it's unfairness to this compare. (Different training data, I can not find the dataset use in original paper )

git clone git@github.com:AceCoooool/NLDF-pytorch.git cd NLDF-pytorch/

Note: the original paper use other datasets.

Download the ECSSD dataset.

bash download.sh

cd tools/ python extract_vgg.py cd ..

python demo.py --demo_img='your_picture' --trained_model='pre_trained pth' --cuda=True

Note:

1. default choose: download and copy the pretrained model to weights directory.
2. a demo picture is in png/demo.jpg

python main.py --mode='train' --train_path='you_data' --label_path='you_label' --batch_size=8 --visdom=True --area=True

Note:

1. --area=True, --boundary=True area and boundary Dice IOU (default: --area=True --boundary=False)
2. --val=True add the validation (but your need to add the --val_path and --val_label)
3. you_data, you_label means your training data root. (connect to the step 2)

python main.py --mode='test', --test_path='you_data' --test_label='your_label' --batch_size=1 --model='your_trained_model'

Note:

1. use the same evaluation (this is a reproduction from original achievement)

1. The boundary Dice IOU may cause inf，it is better to use area Dice IOU.

Maybe, it is better to add Batch Normalization.