Behavioral experiments have shown that the human visual system makes extensiveuse of scene contextual information for facilitating object detection andrecognition. However, the question of how to formally model such contextualinfluences is still largely open. Most of research on computer vision focuseson the detection of single object classes (like faces or cars) considering thebackground as a collection of distractors. Similarly, current computationalmodels of visual attention focus on bottom-up information, based on localfeatures, and ignore top-down information given by the scene context. In thiswork, I consider the problem of detecting and recognizing simultaneously alarge number of different classes of objects in context. By studying the moregeneral problem of multiclass object detection it is possible to devise newalgorithms, more efficient and robust than the ones devoted to detecting oneobject class at a time. I will present an algorithm that learns to recognizemultiple objects jointly, exploiting contextual relationships between objects.I will describe how scene information can then be used to modulate the saliencyof image regions early during the visual processing in order to provide anefficient short cut for object detection and recognition. Finally, I willcompare the performance of a saliency-based model and a model incorporatingcontextual information in predicting eye movements by observers performing a�people search� task.
�
