Abstract
Our vision is unsurpassed by machines because we use a sophisticated object representation. This representation is unlike the retinal image: on the one hand, two out-of-phase checkerboards, maximally different in image pixels, appear perceptually similar. On the other hand, two faces, similar in their image pixels, appear perceptually distinct. What is then the nature of perceptual space? Are there principles governing its organization? To address these questions, we have been using visual search to characterize similarity relations between objects. Compared to the classical approach of asking subjects to provide subjective dissimilarity ratings, visual search has the advantage that it is a natural, objective task where performance is implicitly related to similarity: the time taken to find a target in visual search depends directly on the similarity between the target and distracters. I will summarize a line of research from our laboratory indicative of a surprising linearity governing perceptual space. In the first study, we found that search time is inversely proportional to the feature difference between the target and distracters. The reciprocal of search time is therefore linear and interestingly, it behaved like a mathematical distance metric. It also has a straightforward interpretation as a saliency signal that drives visual search (Arun, 2012). In a second study, complex searches involving multiple distracters were explained by a linear sum of pair-wise dissimilarities measured from simpler searches involving homogeneous distracters (Vighneshvel & Arun, 2013). In a third study, dissimilarities between objects differing in multiple features were found to combine linearly. Even integral features such as the length and width of a rectangle combined linearly upon including aspect ratio as an additional feature (Pramod & Arun, 2014). In a fourth study, distances between multi-part objects were explained as a linear sum of part dissimilarities (Pramod & Arun, submitted to VSS 2015).
Meeting abstract presented at VSS 2015