DSA^2 F: Deep RGB-D Saliency Detection with Depth-Sensitive Attention and Automatic Multi-Modal Fusion (CVPR'2021, Oral)

This repo is the official implementation of "DSA^2 F: Deep RGB-D Saliency Detection with Depth-Sensitive Attention and Automatic Multi-Modal Fusion"

by Peng Sun, Wenhu Zhang, Huanyu Wang, Songyuan Li, and Xi Li.

Prerequisites

Ubuntu 18
PyTorch 1.7.0
CUDA 10.1
Cudnn 7.5.1
Python 3.7
Numpy 1.17.3

Training

Please see launch_train.sh and launch_pretrain.sh for imagenet pretraining and sod training, respectively.

Testing

Please see launch_test.sh for testing on the sod benchmarks.

Main Results

Dataset	E_r	S_λ^mean	F_β^mean	M
DUT-RGBD	0.950	0.921	0.926	0.030
NJUD	0.923	0.903	0.901	0.039
NLPR	0.950	0.918	0.897	0.024
SSD	0.904	0.876	0.852	0.045
STEREO	0.933	0.904	0.898	0.036
LFSD	0.923	0.882	0.882	0.054
RGBD135	0.962	0.920	0.896	0.021

Saliency maps and Evaluation

All of the saliency maps mentioned in the paper are available on GoogleDrive or BaiduYun(code:juc2).

You can use the toolbox provided by jiwei0921 for evaluation.

Additionally, we also provide the saliency maps of the STERE-1000 and SIP dataset on BaiduYun(code:qxfw) for easy comparison.

Dataset	E_r	S_λ^mean	F_β^mean	M
STERE-1000	0.928	0.897	0.895	0.038
SIP	0.908	0.861	0.868	0.057

Citation

@inproceedings{Sun2021DeepRS,
  title={Deep RGB-D Saliency Detection with Depth-Sensitive Attention and Automatic Multi-Modal Fusion},
  author={P. Sun and Wenhu Zhang and Huanyu Wang and Songyuan Li and Xi Li},
  journal={IEEE Conf. Comput. Vis. Pattern Recog.},
  year={2021}
}

License

The code is released under MIT License (see LICENSE file for details).

foreverfei/DSA2F