This project aims to recreate the results of Zhang et al 2008 by implementing a Bayesian framework to compute visual saliency based on natural image statistics. Additionally, we include a prior over the pixels to increase accuracy.

The image below shows an input image overlayed with the fixation maps (red dots). Next to it you will find just the fixation map preprocessed with a Gaussian kernel, which we will refer to as the ground-truth data. The rest is a comparison between our method and a similar method.

Saliency can be seen as a proxy for visual attention. The general theory is that image features that are highly salient (popping out) are also features are likely to be attended to. This is however a purely bottom-up approach to attention and does not address the more difficult problem of how to model top-down attention mechanisms.

In order to estimate the saliency of a pixel Zhang et al. have derived a Bayesian framework for saliency. You can read more about this framework in the linked paper. In short we do:

1. Calculate ICA filters from a set of natural images.
2. Compute responses by convolving filters with another set of images.
3. Fit a generalized Gaussian distribution to responses of each filter.
4. Compute saliency using parameters from the Gaussian.
5. Convolve saliency map with a normal distribution centered in the middle of the image (prior).

The main.py file will compute the ICA components and the parameters of the distributions. Make sure that eizaburo-doi-kyoto_natim-c2015ff is located in the same directory along with CAT2000_train.

Then run the main file with the id of the file you want to estimate saliency for. For example:

python main.py 11