Detección de la saliencia visual dinámica: hacia el modelado de la atención

Abstract

The evaluation of the features of an image analysing the eye movements is important for developing computer vision applications, as well as the understanding of how biological systems explore the environment. For this reason, there are numerous models that seek to predict where the human visual attention will be focused when watching an image, i.e. its visual saliency. In this project, a dynamic saliency model has been modified towards the attention modeling. This has been done by adding a face detector and a center-bias to an existing saliency model. We will start explaining some saliency and face detection algorithms, delving into the ones that have been used in order to carry out our objective. Then, the description of the developed model is given, together with the decisions that have been taken. Finally, we included some experiments and its results to evaluate the performance of the implemented system. A budget showing the costs of improving the algorithm is given in chapter 6. Nowadays, most saliency models are focused on the extraction of bottom-up information, like color, contrast or motion, to predict the eye fixations of an observer. Nevertheless, human visual attention focuses on high-level features of the image, which provide relevant information in order to understand the scene –top-down attention–. A small amount of visual saliency algorithms include this top-down attention, so our purpose is to improve the performance of an existing bottom-up saliency model, by adding some high-level features: the tendency of looking to the center of the screen and the fact that we pay more attention to faces. These two top-down cues were included, although the proposed problem was to implement only the existence of faces. In order to build this hybrid model, we are going to start from a dynamic bottom-up saliency algorithm, developed by Carlos Ruiz in his Final Year Project (in Spanish, Trabajo Fin de Grado, TFG) –explained in section 3.1–. So, the main objective is to improve its performance. The obtained results and the comparison between both models can be seen in section 4.4. On the other hand, enlarging the eye tracking database was proposed as future work in his TFG, so other objective is to accomplish this task. Although a large database formed by 60 videos with eye fixations data was found, only 8 of them have been used due to the time restrictions, as it is detailed in section 4.1.