Comparing Speed-of-Sight studies using rendered vs. natural images

Abstract

Viewpoint invariant object recognition is both an essential capability of biological vision and a key goal of computer vision systems. A critical parameter in biological vision is the amount of time required to recognize an object. This time scale yields information about the algorithm used by the brain to detect objects. Studies that probe this time scale (speed-of-sight studies) performed with natural images are often limited because image content is determined by the photographer. These studies rarely contain systematic variations of scale, orientation and position of the target object within the image. Semi-realistic three-dimensional rendering of objects and scenes enables more systematic studies, allowing the isolation of specific parameters important for object recognition. To date, a computer vision algorithm that can distinguish between cats and dogs has yet to be developed and the specific cortical mechanisms that enable biological visual systems to make such distinctions are unknown. We perform a systematic speed-of-sight study as a step towards developing such an algorithm by enabling a better understanding of the corresponding biological processing strategies. In our study, participants are given the task of reporting whether or not a cat is present in an image (‘cat / no cat’ task). The object image is displayed briefly, followed by a mask image. As a mask, we use images of dogs as well as 1/f noise. We perform studies with natural images and with rendered images and compare the results.