Abstract
Purpose: :
Watson, Barlow, & Robson (1983) rated spatial–temporal stimuli in terms of the least contrast energy necessary for detection and found a drifting Gabor to be the best stimulus. But, is a Gabor the best static image? And what stimulus does the visual system seem to be tuned for when contrast sensitivity is accounted for?
Methods: :
The Modelfest data set has 43 different images whose detection thresholds have been measured for 16 observers from several different laboratories. Watson (2000 Optics Express) and Watson and Ahumada (2005 Journal of Vision) have fit a range of models to these thresholds that take into account the contrast sensitivity function. Stimuli whose thresholds are below these model predictions are those for which the rest of the visual system is performing well.
Results: :
When the thresholds for the different stimuli are plotted in contrast energy, the two best stimuli are a one–octave–bandwidth Gabor patch at 4 cycles per degree (similar to the Watson et. al (1983) result) and (the best stimulus) a Gaussian blob with a standard deviation of 2.1 arc min. When contrast sensitivity is taken into account, the best Gaussian blob is the smallest one (1.0 arc min) and the overall best stimulus is the smallest Gabor, the one–octave–bandwidth patch at 16 cycles per degree. The detectability of these small images can be ascribed to the fact that spatial summation in the fovea is much weaker than energy summation. When summation is weakened by allowing the exponent to rise above 2, the Gabors and the Gaussians are better predicted and the best stimulus is a Gaussian–windowed line with a width of 0.5 arc min and a length standard deviation of 0.5 deg. The advantage for the line remains when Gabor channels are introduced into the model. Note that there is just one line in the 43 stimuli so that none of the line parameters have been varied.
Conclusions: :
For the Modelfest images, the most detectable image in contrast energy is a Gabor or a small Gaussian. With contrast sensitivity compensation, the most detectable images are the smallest Gabors and Gaussians. When spatial summation is considered, the line is the most detectable. This advantage for lines is supported by two recent results from our lab. We have reported that blurred cracks are poorly detected relative to unblurred cracks and that dipoles are poorly detected relative to lines (Ahumada & Beard, 2005 OSA Fall Vision Meeting). It appears we should have said it the other way.
Keywords: detection • pattern vision • perception