Depth Quality-aware Selective Saliency Fusion for RGB-D Image Salient Object Detection

Prerequisites

| Caffe | CUDA10 | CUDNN7.5 | Matlab2016b |

Usage

1. Clone this code by git clone https://github.com/XueHaoWang-Beijing/DQSF.git --recursive,
   assume your source code directory is $DQSF/
2. Calculate our proposed depth quality-aware features (or download them directly from Google drive For Testing set For Training set or Baiduyun PW:zeht ).

• For RQ and SM features, run the code ./DQSF/Features/RDQ.m.
• For SMM features, run the code ./DQSF/Features/SMM_Network/Test/tesDemo.m and ./DQSF/Features/SMM_Network/Test/tesDemor.m to generate the RGBD and RGBDrand predictions.
  Then run the code ./DQSF/Features/SSMG.m to calculate the SMM features.

Training

1. Download training data(Google drive or Baiduyun PW:4w6j), and extract it to ./DQSF/Dataset/
2. Download initial model and put it into ./DQSF/Network/Train/Model/
3. Start to train with sh ./DQSF/Network/Train/finetune.sh.

Testing

1. Download pretrained model and RGBD datasets into the ./DQSF/Network/Test/model/ and ./DQSF/Network/Test/data/ separately
2. Generate saliency maps by run the code ./DQSF/Network/Test/tesDemo.m