Awesome

Data-Level Recombination and Lightweight Fusion Scheme for RGB-D Salient Object Detection

This repository contains code for paper "Data-Level Recombination and Lightweight Fusion Scheme for RGB-D Salient Object Detection".

• The "Ours" folder contains the source code of our network, which utilizes original RGB and depth maps to calculate the saliency maps.
• The "Ours+" folder contains the source code of our improved network, which replaces the original depth maps with the saliency predictions generated above.

Results

Google Drive

Ours: | All | DES | DUT | LFSD | NJU2K | NLPR | SIP | SSD | STERE |

Ours+: | All | DES+ | DUT+ | LFSD+ | NJU2K+ | NLPR+ | SIP+ | SSD+ | STERE+ |

BaiduYunPan

Ours: | All(gw7w) | DES(d9xt) | DUT(bppe) | LFSD(ibij) | NJU2K(nuh6) | NLPR(bk48) | SIP(3qiy) | SSD(2jbt) | STERE(rbhb) |

Ours+: | All(sjzf) | DES+(vshi) | DUT+(ziw3) | LFSD+(v4ey) | NJU2K+(6w4n) | NLPR+(2cns) | SIP+(tr7c) | SSD+(vy9r) | STERE+(39hc) |

The evaluation can be seen in folder '/Result/Evaluation_Curves/'

Prerequisites

| caffe-master | CUDA10 | CUDNN7.5 | Matlab2016b |

Testing

1. Download Testing sets GoogleDrive BaiduYun(5nga) and extract it to ./

2. Download our pretrained model GoogleDrive BaiduYun(ywsx) and store them to './model/'

• "Ours.caffemodel" :For the results
• "Ours+.caffemodel" :For the final improved results

3. Run the test demo

• Firstly, run "test.m" in "Ours" folder to generate the saliency maps.
• Secondly, run "test+.m" in "Ours+" folder to obtain the improved results.

Training

1. Download training data GoogleDrive BaiduYun(f8hk) and extract it to ./Dataset/Train/
2. Download initial VGG16 model GoogleDrive BaiduYun(ftss) and put it into ./Model/
3. Start to train our network with sh ./ours/finetune.sh.
4. Calculate the saliency predictions refer to "Testing 3.1".
5. Start to train our improved network with sh ./ours+/finetune.sh.