Evaluation of Selective Attention under Similarity Transforms

Computational selective attention systems have mostly been developed as models of human attention, and they have been evaluated on that basis. Now, however, they are being used as front ends to object recognition systems, and in particular to appearance-based recognition systems. As such, they need to be evaluated by other criteria. A common goal for object recognition systms in invariance to 2D similarity transformations (i.e. in-plane translations, rotations, reflections and scales). This implies that attention systems used as front ends should also be invariant to similarity transforms. This paper evaluates the Neuromorphic Vision Toolkit (NVT), a well known and publicly available selective attention system, and finds it to be highly sensitive to 2D similarity transforms. Further investigation, however, suggests that this sensitivity is an artifact of the publicly available implementation, and not of the neuromorphic principles it is based on. Therefore we have developed a new system, called SAFE (Selective Attention as a Front End), that is conceptually similar to NVT. However, SAFE is largely invariant to 2D similarity transformations of the source image and selects scales as well as spatial locations for fixations, implementing a combined “zoom-spotlight” model of attention.