TJUMMG / DIPONet

DIPONet: Dual-Information Progressive Optimization Network for Salient Object Detection

DIPONet: Dual-Information Progressive Optimization Network for Salient Object Detection.

This is a PyTorch implementation of our proposed DIPONet for SOD. It has been accepted by DIGITAL SIGNAL PROCESSING.

Prerequisites

• Pytorch 0.4.1+
• torchvision
• scipy 1.2.1
• opencv-python 4.4.0.44

Usage

1. Clone the repository

git clone https://github.com/TJUMMG/DIPONet.git
cd DIPONet/

2. Download the datasets

Download the training dataset and unzip it into ./data folder.

• [DUTS-TR]

Download the testing datasets and unzip them into ./data folder.

• [DUTS-TE]
• [ECSSD]
• [HKU-IS]
• [PASCALS]
• [SOD]
• [DUTOMRON]

For edge label, you can calculate them by using "get_edge_data.py". And you should put edge labels and sal labels together. If you can't find these public SOD datasets, please concat us via: [e-mail].

3. Download the pre-trained models for backbone (VGG16 and ResNet50)

Download the following pre-trained models [BaiDuYun]：https://pan.baidu.com/s/1qc6zgWf3aDAre6Ey_KQGlw 提取码：4o44 into ./pretrained folder.

4. Train

1. Set the image_root and gt_root path in train.py correctly.

2. We demo using ResNet-50 and VGG-16 as network backbone and train with a initial lr of 2e-5 for 26 epoches, which is divided by 2 after 14 and 22 epochs. The input size is 448 * 448 and the batch size is 5.

3. We demo joint training with edge, you should put edge labels with saliency labels together.

4. After training the result model will be stored under ./trainresults/resnet or ./trainresults/vgg folder.

5. Test

1. Set the test_root and dataset and model.load_state_dict(..) path in test.py correctly.

2. The input size for testing is 448 * 448.

3. We demo joint training with edge, you can get edge results and saliency results.

4. After testing, the result images will be stored under ./testresults/DIPONet_ResNet or ./testresults/DIPONet_VGG folder.

6. Pre-trained models, pre-computed results and evaluation tools

1. The DIPONet model trained by authors [BaiDuYun]：https://pan.baidu.com/s/1C-k2gepxcpPbHR4QEjbD6A 提取码：frf6
2. we provide Saliency maps calculated by ourselves [BaiDuYun]: https://pan.baidu.com/s/1KLJxZzALrUflSj2NI-mcAg 提取码：0jdg
3. All the evaluation results are calculated by using https://github.com/ArcherFMY/sal_eval_toolbox.

7. Contact

If you have any questions, feel free to contact us via: [e-mail] or [e-mail].