Object recognition in egocentric videos with saliency-based non uniform sampling and variable resolution space for features selection

Since recently a new video content is massively coming into practice: the egocentric videos recorded by body-worn cameras. In the context of this work which is the behavioral study patients with Alzheimer disease, this kind of video content allows for a close-up view of instrumental activities of daily living (IADL). In parallel, automatic extraction of visually salient areas from this kind of video content is a strong research direction since it brings the focus of attention on interacted objects (manipulated, observed) during IADLs. Recognition of manipulated objects is a key cue for an automatic activity assessment. In this work we describe our approach for object recognition using visual saliency modeling. We build our model on the well-known BoW paradigm, and propose a new approach to add saliency maps in order to improve the spatial precision of the baseline approach. Finally we use a non-linear classifier to detect the presence of a category in the image. In this research, the contribution of saliency is twofold: • It controls how and where circular local patches are sampled in an image for descriptor computation. • It controls the spatial resolution at which the features are computed. Our aim is to emulate the retina in the Human Visual System (HVS) where cells in charge of foveal and peripheral vision work at different spatial resolutions.