Abstract
The same object can look different to an observer because of changes in viewpoint, lighting, and other parameters. This makes object recognition that is invariant to viewing condition a staggering challenge. It has been suggested that spatiotemporal association can support invariant learning; combining information gained during distinct encounters can improve one’s estimates of an object’s true features. But spatiotemporal association is insufficient; one needs to know whether spatiotemporally close encounters include the same object. We hypothesized that spatiotemporal association should therefore be constrained by the rules of object persistence, the object physics thought to make up core knowledge in infants and known to guide online perception of token identity. We tested this hypothesis with an incidental-encoding paradigm followed by a recognition test. During each encoding trial a single object appeared twice, via an apparent motion manipulation, in a way that made it look like encounters with two distinct individuals (discontinuous motion) or that made it look like two encounters with the same individual (continuous motion). Participants then judged test objects each as old, similar, or new. Signal detection measures of recognition were used to compare performance as a function of motion continuity. In Experiment 1, encoding involved encounters with an object embedded each time in independent noise. In Experiment 2, encoding occurred without noise, but test images were embedded in varying degrees of noise. And in Experiment 3, each encoding encounter included the object oriented differently, with old objects at test appearing at a third, never-before-seen orientation. (Additional control and extension experiments were run as well). In all three experiments, objects encountered through continuous motion were recognized significantly better than objects encountered through discontinuous motion. These results demonstrate how invariant object learning is supported by constraints from object physics that control the online perception of token identity.
Meeting abstract presented at VSS 2015