Computational Models of Perceptual Space : From Simple Features to Complex Shapes

Abstract

Dissimilarity plays a very important role in object recognition. But, finding perceptual dissimilarity between objects is non-trivial as it is not equivalent to the pixel dissimilarity between the objects (For example, two white noise images appear very similar even when they have different intensity values at every corresponding pixel). However, visual search allows us to reliably measure perceptual dissimilarity between a pair of objects. When the target object is dissimilar to the distracter, visual search becomes easy and it will be difficult otherwise. Even though we can measure perceptual dissimilarity between objects, we still do not understand either the underlying mechanisms or the visual features involved in the computation of dissimilarities. For this thesis, I have explored perceptual dissimilarity in two studies – by looking at known simple features and understanding how they combine, and using computational models to understand or discover complex features. In the first study, we looked at how dissimilarity between two simple objects with known features can be predicted using dissimilarities between individual features. Specifically, we investigated how search for targets differing in multiple features (intensity, length, orientation) from the distracters is related to searches for targets differing in each of the individual features. We found that multiple feature dissimilarities could be predicted as a linear combination of individual feature dissimilarities. Also, we demonstrated for the first time that Aspect ratio of the object emerges as a novel feature in visual search. This work has been published in the Journal of Vision (Pramod & Arun, 2014). Having established in the first study that simple features combine linearly, we devised a second study to investigate dissimilarities in complex shapes. Since it is known that shape is one of the salient and complex features in object representation, we chose silhouettes of animals and abstract objects to explore the nature of dissimilarity computations. We conducted visual search using pairs of these silhouettes on humans to get an estimate of perceptual dissimilarity. We then used various computational models of shape representation (like Fourier Descriptors, Curvature Scale Space, HMAX model etc) to see how well they can predict the observed dissimilarities. We found that many of these computational models were able to predict the perceptual dissimilarities of a large number of object pairs. However, we also observed many cases where computational models failed to predict perceptual dissimilarities. The manuscript related to this study is under preparation.