Purchase this article with an account.
Ljiljana Velisavljevic, James H. Elder; What do we see in a glance?. Journal of Vision 2002;2(7):493. doi: 10.1167/2.7.493.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
OBJECTIVE. Prior studies have suggested that rapid visual recognition proceeds from course-to-fine spatial scales. Here we examine whether this holds for local recognition tasks and to what degree this processing can be understood in the spatial frequency domain. METHOD. We adapted the priming method of Tjan et al. (1999). In the test phase, subjects were briefly presented with colour images of natural scenes, each divided into 64 100×100 pixel subimages, and preceded and followed by 1/f colour noise masks. In the first condition, test images were presented intact; in a second condition the 64 subimages were randomly scrambled. After each test image, subjects were shown 2 subimages and asked to indicate which was drawn from the test image. QUEST was used to estimate the test image stimulus duration for 82% correct performance. RESULTS. Expt.1 compared subimage recognition for intact and scrambled test images. Scrambling was found to increase the stimulus duration required for local recognition. Thus global information plays a major role in fast scene recognition, even for local scene components. In Expt.2, test images were masked in a checkerboard fashion, so that only alternate subimages were visible. We found that recognition was independent of whether the test subimage had been visible or masked. Thus fast recognition appears to depend upon local information only insofar as it contributes to the global statistics of the scene. Expt.3 compared recognition for normal images and images inverted in both orientation and contrast. Inversion was found to significantly degrade recognition performance. Moreover, for inverted images, the effect of scrambling disappeared. Since the spatial frequency spectra of normal and inverted images are essentially identical, these results suggest that coarse-to-fine analysis in scene recognition cannot be understood in the spatial frequency domain. Rather, fast scene recognition appears to depend upon more semantic global information.
This PDF is available to Subscribers Only