This repository is the implementation of 'Video Saliency Prediction using Spatiotemporal Residual Attentive Networks (TIP2019)'.
Qiuxia Lai, Wenguan Wang, Hanqiu Sun, Jianbing Shen
- CentOS-7
- Python 3.5.2
- Tensorflow 1.11.0
- Keras 2.2.4
- CUDA 9.0
- CUDNN 7.5.0
Prediction results on DHF1K, Hollywood-2, UCF sports, and DIEM can be downloaded from:
Google Drive: Pending ...
Baidu Disk: This link password:dizw
The optical flows are generated using flownet-2.0. The '.flo' files in the '/flow' folders under the video directory. Please be noted that the optical flow files may take up a considerable amout of storage space. The dataset directory is as follows:
DataSets
|----DHF1K
|---train
|--0001
|----images
|----flow
|----fixation
|----maps
|--0002
|----...
|--...
|---test
|--...
|----Hollowood-2
|---train
|---test
|----UCF sports
|---train
|---test
For Holloyood-2, we further seperate the video sequences into shots according to the ground-truth shot boundaries, and discard the ones that contains less than 10 frames.
For more information about DHF1K, click here. See DHF1K leaderboard here
Google Drive: This link.
Baidu Disk: This link. password:pqil
The testing weights U_td3_res_ds.h5 and UHD_dcross_res_matt_res.h5 are put into '/vap_model' by default. The initialization weights A_resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5 and M_resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5 are put into '/weights' by default.
Please modify the config.py accordingly.
1.Easy testing: run demo_test.py to get result from one video by default.
A video test example from UCF sports can be found in the /DataSets folder. The results would be stored in '/vap_predictions' by default.
2.To get other testing results, prepare the datasets with optical flows, and modify the dataset settings in config.py. Run demo_test.py.
You may also run demo.py after editing the config.py with model_no = 0 or 1 and phase = 'vis'， where 0 is for the feature net, and 1 is for the whole model.
3.To visualize the multi-scale masks or the side outputs, please modify the prediction part of demo.py or demo_test.py.
Our model is trained in tendem. To train the feature net, we initialize it with the weight of resnet-50 pretrained on the ImageNet. Then, we initialize the feature net with the weight in the first step, randomly initialize the remaining part, and train the whole network.
In case that you do not want to train from the feature net, you may directly use the provided weight U\_td3\_res_ds.h5, and begin from step 2.
1.Train the feature net
1)In config.py, set model_no = 0 and phase = 'train'
2)Run demo.py
2.Train the whole net
1)In config.py, set model_no = 1 and phase = 'train'
2)Modify the initilization weight of the feature net in demo.py to the one you obtained in step 1, or leave it as the default one that you've downloaded.
3)Run demo.py
If you use our code in your research work, please consider citing the following papers:
@ARTICLE{lai2019video,
title={Video Saliency Prediction using Spatiotemporal Residual Attentive Networks},
author={Qiuxia Lai and Wenguan Wang and Hanqiu Sun and Jianbing Shen},
journal={IEEE Trans. on Image Processing},
year={2019}
}
@InProceedings{Wang_2018_CVPR,
author = {Wang, Wenguan and Shen, Jianbing and Guo, Fang and Cheng, Ming-Ming and Borji, Ali},
title = {Revisiting Video Saliency: A Large-Scale Benchmark and a New Model},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition},
year = {2018}
}
@ARTICLE{Wang_2019_revisitingVS,
author={W. {Wang} and J. {Shen} and J. {Xie} and M. {Cheng} and H. {Ling} and A. {Borji}},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
title={Revisiting Video Saliency Prediction in the Deep Learning Era},
year={2019},
}
Qiuxia Lai: ashleylqx@gmail.com