RGB-T Salient Object Detection via Excavating and Enhancing CNN Features

Figure.1 The overall architecture of the proposed E2Net model, which is published in Applied Intlligence 🎆. The paper can be downloaded from here[code:NEPU].

1.Requirements

Python v3.6, Pytorch 0.4.0+, Cuda 10.0, TensorboardX 2.0, opencv-python

2.Data Preparation

Download the train data from here[code:NEPU], test data from here[code:NEPU], test_in_train data from here[code:NEPU]. Then put them under the following directory:

-Dataset\   
   -train\  
   -test\ 
       -VT821\
       -VT1000\
       -VT5000_test\
   -test_in_train\

3.Training/Testing & Evaluating

• Training the E2Net

Please download the released code and the data set, then:

run python train.py  

• Testing the E2Net

Please download the trained weights from here[code:NEPU], and put it in './pre' folder, then:

run python test.py  

Then the test maps will be saved to './Salmaps/'

• Evaluate the result maps

You can evaluate the result maps using the tool from here[code:NEPU], thanks for Dengpin Fan.

4.Results

• Qualitative comparison

Figure.2 Qualitative comparison of our proposed method with some SOTA methods.

• Quantitative comparison

Table.1 Quantitative comparison with some SOTA models on three public RGB-T benchmark datasets.

• Salmaps
  The salmaps of the above datasets can be download from here [code:NEPU]

5.Contact

If you have any questions, feel free to contact us via wuranwan2020@sina.com (Ranwan Wu). For more related work, you can also visit wuranwan.