Abstract
Clutter can degrade performance at many visual tasks. Though the degree to which this is true depends on the task, task-independent measures of clutter have proven useful for replacing notions of set-size for natural scenes, and for predicting subjective clutter judgments, search performance, eye movements, and visibility of artifacts. Rosenholtz, Li, & Nakano (2007) proposed several clutter measures: The Feature Congestion measure derives from the suggestion that clutter makes it difficult to add a salient, easy-to-find item; higher feature variance means more clutter. On the other hand, the Subband Entropy clutter measure follows from the intuition (and observations) that the more cluttered an image the more bits one needs to encode it. However, a prominent factor in the degradation in performance due to clutter has to do with peripheral crowding. Peripheral vision pools texture statistics over sizable regions that grow with eccentricity (Balas et al., 2009; Freeman & Simoncelli, 2011). The degradation of performance due to clutter may therefore depend on the cost of that summarization. This intuition connects to the aforementioned clutter measures, since crowding affects ease of search (e.g. saliency), and one might reasonably associate an entropy measure with the cost of summarizing a patch. This work develops a new multifeatured clutter measure. It can be foveated, i.e. for known fixation it depends on eccentricity. Rather than measuring simple subband entropy, it depends on the texture statistics hypothesized to underlie peripheral crowding, such as the joint statistics of magnitude of oriented subbands. This measure has advantages over previous clutter measures; for example, it marks some textures as relatively uncluttered, in line with predictions of texture statistic models of peripheral vision.