Evaluation of selective attention under similarity transformations

摘要

Selective attention systems have generally 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. We argue that to serve as effective front ends for object recognition, selective attention systems should (1) select fixation points in scale as well as position, and (2) be insensitive to 2D similarity transformations of the image (i.e., in-plane translations, rotations, reflections, and scales). This paper evaluates the Neuromorphic Vision Toolkit (NVT), a well-known selective attention system, and finds that it satisfies neither criterion. Further investigation, however, suggests that the sensitivity of NVT to similarity transformation is an artifact of its implementation. We develop a new system, called selective attention as a front end (SAFE), that is based on the same principles as NVT, but selects both the scale and position of fixation points and is largely invariant to 2D similarity transformations.