We propose a Dual Attention Network (DANet) to adaptively integrate local features with their global dependencies based on the self-attention mechanism. And we achieve new state-of-the-art segmentation performance on three challenging scene segmentation datasets, i.e., Cityscapes, PASCAL Context and COCO Stuff-10k dataset.

We train our DANet-101 with only fine annotated data and submit our test results to the official evaluation server.

1. Install pytorch

• The code is tested on python3.6 and official Pytorch@commitfd25a2a, please install PyTorch from source.
• The code is modified from PyTorch-Encoding.

1. Clone the repository:

   git clone https://github.com/junfu1115/DANet.git 
   cd DANet 
   python setup.py install

2. Dataset

• Download the Cityscapes dataset and convert the dataset to 19 categories.
• Please put dataset in folder ./datasets

4 . Evaluation

• Download trained model DANet101 and put it in folder ./danet/cityscapes/model

• Evaluation code is in folder ./danet/cityscapes

• cd danet

• For single scale testing, please run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python test.py --dataset cityscapes --model danet --resume-dir cityscapes/model --base-size 2048 --crop-size 768 --workers 1 --backbone resnet101 --multi-grid --multi-dilation 4 8 16 --eval

• For multi-scale testing, please run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python test.py --dataset cityscapes --model danet --resume-dir cityscapes/model --base-size 2048 --crop-size 1024 --workers 1 --backbone resnet101 --multi-grid --multi-dilation 4 8 16 --eval --multi-scales

• If you want to visualize the result of DAN-101, you can run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python test.py --dataset cityscapes --model danet --resume-dir cityscapes/model --base-size 2048 --crop-size 768 --workers 1 --backbone resnet101 --multi-grid --multi-dilation 4 8 16

1. Evaluation Result:

   The expected scores will show as follows:

   (single scale testing denotes as 'ss' and multiple scale testing denotes as 'ms')

   DANet101 on cityscapes val set (mIoU/pAcc): 79.93/95.97 (ss) and 81.49/96.41 (ms)

2. Training:

• Training code is in folder ./danet/cityscapes
• cd danet

You can reproduce our result by run:

 CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset cityscapes --model  danet --backbone resnet101 --checkname danet101  --base-size 1024 --crop-size 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 2 --multi-grid --multi-dilation 4 8 16

Note that: We adopt multiple losses in end of the network for better training.

If DANet is useful for your research, please consider citing:

@article{fu2018dual,
  title={Dual Attention Network for Scene Segmentation},
  author={Fu, Jun and Liu, Jing and Tian, Haijie, and Fang, Zhiwei, and Lu, Hanqing},
  journal={arXiv preprint arXiv:1809.02983},
  year={2018}
}

Thanks PyTorch-Encoding, especially the Synchronized BN!