Abstract
Classification images have become an important tool for understanding visual processing in tasks limited by noise. However, with the exception of a few studies, the technique is currently limited by requiring that targets and distracters be well cued for location in yes-no or forced-choice psychophysical experiments. Here, we investigate the method in free-localization tasks, where subjects search a single contiguous image for a target and respond by indicating the location where the target is believed to be positioned. Subject responses can be acquired by a mouse or other pointing device, or indirectly by an eye-tracker. Free localization tasks have a number of attractive qualities, including controllable incorporation of free search into detection tasks, higher target contrast, and a more informative subject response that results in fewer images needed to estimate a classification image. The approach we propose involves averaging incorrect localizations after alignment and correcting for the correlation structure of the noise.
We have evaluated the proposed methods using linear filter models for a Gaussian luminance target embedded in Gaussian noise having power-law amplitude spectra with exponents from 0 (white noise) to -3/2. We find the result of overlapping (i.e. dependent) locations is a consistent underestimation in the lowest spatial frequencies in the classification images, which becomes more pronounced as the exponent decreases. Small motor errors in the localization response relative to the size of the target have little effect on the resulting classification images. However, as the motor errors exceed the target size, the classification images show consistent underestimation at high spatial frequencies that is also dependent on the exponent of the noise process. We find approximately a factor of two or more reduction in the number of trials needed to obtain a classification image with comparable signal-to-noise ratio to a yes-no task.