SalAR - Video Saliency

This is a trimmed version of the code. Full code with detailed documentation, and all models, will be available soon.

Dependencies:

Install - if necessary - the required dependencies:

• Python (tested with 3.7.3, conda 4.6.14)
• PyTorch (tested with PyTorch 1.1.0, CUDA 9.2, scipy 1.2.1)
• Other python dependencies: numpy, scipy, matplotlib, opencv-python (cv2)
• CUDA is required.

Quick Inference

We have added a test sequence from the UCF-Sports dataset in the dataloaders folder.

1. Run python test_sequence.py.

Inference

1. Download Pre-trained UCF-Model.
2. Download the model given for UCF-Sports dataset. Place it in models folder.
3. Download the datasets and place them under dataloaders. They can be downloaded from https://github.com/wenguanwang/DHF1K. The dataset directory should look like the following:

└── dataloaders
│   ├── ucf
│   │   ├── training
│   │   	└── sequence_name
│   │   		├── images
│   │   		├── fixation
│   │   		└── maps
│   │   └── testing
│   │   	└── sequence_name
│   │   		├── images
│   │   		├── fixation
│   │   		└── maps
│   ├── hollywood
│   │   ├── training
│   │   	└── sequence_name
│   │   		├── images
│   │   		├── fixation
│   │   		└── maps
│   │   └── testing
│   │   	└── sequence_name
│   │   		├── images
│   │   		├── fixation
│   │   		└── maps
│   └── dhf1k
│   │   ├── training
│   │   	└── sequence_name
│   │   		├── images
│   │   		├── fixation
│   │   		└── maps
│   │   └── testing
│   │   	└── sequence_name
│   │   		├── images
│   │   		├── fixation
│   │   		└── maps

2. For DHF1K, please put the validation set files in the testing directory (the directory is named testing to streamline code).
3. Run python inference_with_metrics.py.