How does your brain decide what a salient object is

First thought on the problem of salience: hmmm, what do the computational folk make of all of this. Are there any algorithms around for spotting salient objects that seem to mimic how meat machines do it?

Chosen at random: a paper by Liu et al (2007). (And they begin with the William James quote: “Everyone knows what attention is…”.) They argue that “visual attention is in general unconsciously driven by low-level stimulus in the scene such as intensity, contrast, and motion.” I buy that. [Note-to-self, where is this going on in the brain?]

Participants were asked to label a load of photos to use as training and test data for the algorithm. The beef of the theory comes from conditional random fields. The problem: does a given pixel belong to the salient object. They plug the following features into the maths: multiscale contrast, \(\chi^2\) distance between histograms of RGB colour, and finally the global distribution of colour (salient objects tend to have colours which are less frequent in the rest of the image).

So their approach combines supervised machine learning and clever selection of imagine features to focus on. Seems to do a good job.

Next on the list… models from closer to the cognitive end of the literature.  Anti pointed me towards Laurent Itti’s lab.  Actually Itti has a nice article in scholarpedia on the topic.


Tie Liu, Jian Sun, Nan-Ning Zheng, Xiaoou Tang and Heung-Yeung Shum. Learning to Detect A Salient Object. CVPR 2007.