Abstract
Surface textures are perceived by computing global summary statistics from ensembles of local visual features, such as the orientation content in the bark of a tree. Most models assume that ensemble perception is achieved by encoding local features and then pooling them across space to compute a global signal. We present evidence that such a pooling process does occur for some stimulus conditions: when asked to report the global orientation of 580 lines with orientations drawn from a Gaussian distribution, errors became gradually smaller as the line orientations became more similar to the average orientation. In addition, confidence judgments revealed that people were aware of how well they were performing in this task, suggesting people have access to the precision of their global orientation representations. However, although it is sometimes reasonable to pool across all local features, there are many cases where such indiscriminate pooling is detrimental to perception. Ensembles are often confounded with noise, such as moss growing on the tree bark, and building accurate ensemble information requires that the signal and noise are segmented from one another rather than pooled. We constructed such stimuli by mixing similarly oriented signal lines with randomly oriented noise lines, and find that global orientation is either perceived with extremely high accuracy, or not at all. Moreover, decreasing the percentage of signal lines led to increases in the probability of completely failing to extract the signal, but global orientation reports remained highly accurate for the remaining trials. These findings suggest that texture segmentation is a discrete, all-or-none process in which global summary statistics are either used to successfully segment one texture from noise, or segmentation fails entirely. Confidence judgments reflected this mixture of accurate segmentation and complete failures to segment, suggesting that our conscious experience of computing ensembles is also all-or-none.