Visual search for objects in a complex visual context: what we wish to see

In this work we propose a saliency based psycho-visual weighting of the BoVW for object recognition. This approach is designed to identify objects related to IADL on videos recorded by a wearable camera. These recording give an egocentric point-of-view on the upcoming action. This point- of-view is also characterized by a complex visual scene with several objects on the frame plan. The human visual system functions is a way to process only the relevant data by considering areas of interest. Based on this idea, we propose a new approach by introducing saliency models to discard irrelevant information in the video frames. Therefore we apply a visual saliency model to weight the image signature within the BoVW framework. Visual saliency is well suited for catching spatio-temporal information related to the observer's attention on the video frame. We also proposed an additional geometric saliency cue that models the anticipation phenomenon observed on subjects watching video content from the wearable camera. The findings show that discarding irrelevant features gives better performances when compared to the baseline method which consider the whole set of features in the images.