|Discriminant Saliency Detection|
Biological vision systems rely on saliency mechanisms to cope with the complexity of visual perception. Rather than sequentially scanning all possible locations of a scene for the occurrence of events of possible interest, these events simply "pop-out" from the scene background. In this way, saliency enables the efficient allocation of perceptual resources and increases the robustness of recognition in highly cluttered environments.
In the computer vision literature, computational definitions of saliency have tended to rely on "universal" saliency principles that result in the detection of generic visual attributes such as edges, contours, corners, parallelism, etc. This lacks both the flexibility and adaptability of biological saliency. This project explores a definition of saliency which is directly grounded on recognition, and argues for saliency detectors that are learned in a discriminant fashion. It has shown that it is possible to learn saliency detectors that have low complexity, are scalable to large scale recognition problems, are compatible with existing models or early biological vision, and can be easily adapted for the recognition of diverse visual concepts.
|Contact:||Dashan Gao, Nuno Vasconcelos|