Personal tools
You are here: Home Publications Learning Saliency. An ICA based model using Bernoulli mixtures
About NIFTi
NIFTi is about human-robot cooperation. About teams of robots and humans doing tasks together, interacting together to try and reach a shared goal. NIFTi looks at how the robot could bear the human in mind. Literally. When determining what to do or say next in human-robot interaction; when, and how. NIFTi puts the human factor into cognitive robots, and human-robot team interaction in particular.   Each year, NIFTi evaluates its systems together with several USAR organizations. Rescue personnel teams up with NIFTi robots to carry out realistic missions, in real-life training areas. 

This site uses Google Analytics to record statistics on site visits - see Legal information.


A. Carbone and F. Pirri (2010)

Learning Saliency. An ICA based model using Bernoulli mixtures

In: Proceedings of BICS, Brain Ispired Cognitive Systems.

In this work, we present a model of both the visual input selection and the gaze orienting behaviour of a human observer undertaking a visual exploration task in a specific scenario. Our input is made of a sequence of gaze-tracked fixations acquired from a custom designed wearable device. We aim at characterising the sta- tistical properties and regularities of the selected visual input. While the structure of the visual context is specified as a linear combination of basis functions, which are independent hence uncorrelated, we show how low level features characterising a scan-path of fixations can be obtained by hidden correlations to the context. Sam- ples from human observers are collected both in free-viewing tasks, in the specified visual scene. These scan-paths show important and interesting dependency from the context. We show that a scan-path, given a database of a visual context, can be suit- ably induced by a system of filters that can be learned by a two stages model: the independent component analysis (ICA) to gather low level features and a mixtures of Bernoulli distributions identifying the hidden dependencies. Finally these two stages are used to build the cascade of filters.