Combining Attention and Temporal Coherence Models: Learning Invariances from Cluttered Scenes

The aim of the present thesis is to model object recognition and spatial visual attention in a single frame. The direction of gaze can serve as a correlate for the alignment of spatial attention, but in conventional laboratory experiments, the subject is confined and the concentration is on eye-in-head movements. We solve this problem by using a mobile head-mounted eye-tracking system EyeSeeCam. In the first part, the camera is analyzed in more detail because of an error of parallax and an endurance test is performed. At the end of my thesis, there is still room for improvement in the functionality. In the second part, we try the increase the understanding of the action of selective visual attention. We quantify how complex cells, as they are known from the primary visual cortex, adapt to the statistics of natural stimuli. A distance-dependent decorrelation factor is integrated into an existing programme to check if orientation columns develop. For new results, further simulations and evaluations are necessary.