tjummg / absnet

ABSNet: Aesthetics-Based Saliency Networkusing Multi-Task Convolutional Network(IEEE Signal Processing Letters2020)

ABSNet

• ABSNet: Aesthetics-Based Saliency Network using Multi-Task Convolutional Network(IEEE Signal Processing Letters2020)
• @ARTICLE{ABSNet,
  author={Jing Liu, Jincheng Lv, Min Yuan, Jing Zhang, and Yuting Su},
  journal={IEEE Signal Processing Letters},
  title={Aesthetics-Based Saliency Network using Multi-Task Convolutional Network},
  year={2020},
  volume={},
  number={},
  pages={},
  doi={10.1109/LSP.2020.3035065} }

This is a Tensorflow implementation of IEEE Signal Processing Letters2020.

Requisites

• opencv-python 4.1.1
• numpy 1.16.1
• tenserflow 1.14.0
• tensorboard 1.14.0
• keras 2.2.2
• scipy 1.2.1

Usage

1. Clone the repository

2. Download the datasets

Download the following datasets and unzip them into data folder.

• SALICON dataset.
• MIT300 dataset.
• MIT1003 dataset.

3. Download the pre-trained models for backbone

Download pre-trained model BaiduYun (pwd: 4jk1).
This pre-trained model is for multi-task dalited resnet(MT-DRN), you should set 'weightfile' correctly.

4. Test our model

• For testing sal branch, you should set config.yaml correctly and load pretrained pkl correctly, and then run main_sal_drn.py.

5. Download code for metrics, evaluation, and more.

https://github.com/cvzoya/saliency

6. Test Results of MT-DRN on SALICON Validation/TEST Set and MIT1003 DataSet.

• SALICON Validation Set: Link：https://pan.baidu.com/s/1dPkwMpotQQhSQql4r42TFQ pwd: oxgw
• SALICON Test Set: Link: https://pan.baidu.com/s/1FygM6C1Su_6bcCivBW3QQw pwd: 6yqq
• MIT1003 DataSet: Link: https://pan.baidu.com/s/1z-q9BLrRBql2AZUUS_5mnA pwd: cwgx

7. Some Comparsions in our paper.