DSLRDNet

This is the official implementation for the paper "Addressing Multiple Salient Object Detection via Dual-Space Long-Range Dependencies", accepted by the Journal of Computer Vision and Image Understanding (CVIU 2023).

Prerequisites:

1. Pytorch 1.2.0
2. Opencv 2.4.5
3. TensorboardX

For training:

1. Download the DUTS-TR (Google Drive) training dataset.
2. Download the initial pratrained VGG/ResNet (Google Drive) model.
3. Change the training data path in dataset.py.
4. Change the training settings in solver.py and run.py
5. Start to train with python3 run.py --mode train

For testing:

1. Download the pretrained models (Google Drive).
2. Change the data path in dataset.py
3. Change the test settings in run.py.
4. Generate saliency maps with python3 run.py --mode test --sal_mode m, where 'm' demonstrates the MSOD dataset.
5. We use the public open source evaluation code. (https://github.com/weijun88/F3Net)

Datasets and results:

MSOD dataset || Generated Saliency Maps (Google Drive)

Citing DSLRDNet:

@article{deng2023addressing,
  title={Addressing multiple salient object detection via dual-space long-range dependencies},
  author={Deng, Bowen and French, Andrew P and Pound, Michael P},
  journal={Computer Vision and Image Understanding},
  volume={235},
  pages={103776},
  year={2023},
  publisher={Elsevier}
}