The Role of Components in Recognition across Changes of View

We examined the extent to which recognition of rotated objects is based on complex features, such as differently shaped components, when distractor objects may project images that are very similar to the studied images. The stimuli were unfamiliar computer-generated objects composed of four long thin components connected end to end. Each object was matched to a distractor object that had the same components arranged in a different order. The 3-D structure of the matched objects was otherwise the same—the lengths of the components in corresponding positions and the connection angles between them were identical. Subjects studied six target objects and then performed a 2AFC recognition task with each target and its matched distractor. Either the target, the distractor, or both were rotated ±40° about the vertical axis from the studied viewpoint. As expected, subjects chose the target object most often when it was not rotated (77.2% and 69.4% correct when the distractor was and was not rotated). However, subjects chose the target object 64.4% of the time when it was rotated and the distractor was not. In this case, the image of the distractor was more similar to the studied image than was the image of the target. Recognition of these objects was not based on pixel-based image similarity among studied and tested objects, but on a more general view-based similarity that takes into account a multitude of features, some more stable over changes of view than others. We discuss computational models that also rely on multiple view-based features.