Abstract
We have previously shown that high-level material categories can be recognized in complex, real world images in exposures as brief as 40 msec (Sharan et al., 2009). This performance is comparable to recent results in rapid object and scene categorization (Bacon-Mace et al., 2005; Greene & Oliva, 2009). We now have studied the speed of object categorization versus material categorization more directly. We collected a database of photographs containing two classes of objects (gloves vs. handbags) that were made from two classes of material (leather vs. fabric). There were a total of 200 images, with 50 examples for each combination of object and material. In a 2AFC task, subjects viewed one image at a time, and categorized each image. For 100 images, the task was object categorization; for the other 100 images it was material categorization. We also ran two baseline RT tasks: telling a blue disc from a red disc, and telling a left tilting line from a right tilting line. We then compared object RT's with Material RT's. As expected, the baseline tasks were the fastest, with a median RT of 390 msec. The object task was slower, but the median RT was only about 50 msec slower than baseline. The material task was slower still, with a median RT more than 150 msec slower than baseline. Since material perception is a complex task (for which we have no good models), 150 msec seems reasonably fast. However, 50 msec is even faster. In a task like this, the precise ratio of the incremental RT's has limited meaning, since one can bias the results by choosing images that are easy or hard on one of the dimensions. However, the 3:1 ratio of incremental RT's for these two tasks is hard to dismiss, and provides clear evidence that object recognition is a faster process than material recognition.