OpenSALICON is an open source implementation of the SALICON saliency model cited below.

Huang, X., Shen, C., Boix, X., & Zhao, Q. (2015). SALICON: Reducing the Semantic Gap in Saliency Prediction by Adapting Deep Neural Networks. In Proceedings of the IEEE International Conference on Computer Vision (pp. 262-270).

A technical report describing our implementation is provided (salicon.pdf). If you find OpenSALICON useful in your research, please cite our technical report:

@TECHREPORT {christopherleethomas2016,
    author      = {Christopher Lee Thomas},
    title       = {OpenSalicon: An Open Source Implementation of the Salicon Saliency Model},
    institution = {University of Pittsburgh},
    year        = {2016},
    number      = {TR-2016-02}
}

1. Requirements for Caffe and pycaffe (see: Caffe installation instructions)

Note: Caffe must be built with support for Python layers!

# In your Makefile.config, make sure to have this line uncommented
WITH_PYTHON_LAYER := 1

2. Python packages you might not have: numpy, PIL, matplotlib, scipy

1. Edit the Salicon.py file to include the path to your caffe python install where indicated (i.e. update the line below to the path containing your caffe installation):

       sys.path.insert(0, 'caffe/install/python') # PATH TO CAFFE PYTHON INSTALL

   Additionally, depending on whether you are using a GPU or not, you may need to edit the following two lines:

   caffe.set_mode_gpu()
   caffe.set_device(1)

   If you do not have a CPU, be sure to remove the set_device(1) line and change the mode selection to caffe.set_mode_cpu(). Depending on your setup, you may need to change the device ID from 1 to 0 or some other number depending on your hardware configuration.

2. Download the caffemodel files using this link. Place the caffemodel files into the same directory as the code (unless you manually modify the paths to point to the directory containing the caffemodel files).

3. The following commands should allow you to compute the saliency map for an image using the pretrained model. IMPORTANT: When using on your own data, ensure that your data is in 256 RGB format. If it is not, you will need to manually do some pre-processing first.

4. Run python and execute the following commands:

   from Salicon import Salicon
   sal = Salicon()
   map = sal.compute_saliency('face.jpg')
   # map contains saliency map in double format.

   

5. You can then perform thresholding (if you prefer) on the output or use it directly.

Use the finetune_salicon.py file to train your own model. See our technical report for more details on how to do this. We provide two solvers solver.prototxt and solver_new.prototxt for this purpose. Solver_new attempts to use ADADELTA to adjust the learning rate dynamically. finetune_salicon.prototxt provides a model definition file for our implementation in which you can adjust learning rates per layer and more customization. Our finetune_salicon.py file provides the basic functionality of reading in the input images and fixation maps, initializing the solver, loading the data in batches, solving the model, and saving the results. You may need to adjust the solver to your particular dataset.

OpenSALICON performs on-par with the results produced by the official SALICON demo website when compared using AUC scores to ground truth human fixations. If you have any questions, discover any bugs, or make improvements to this code, please feel free to create an issue or push request.