ausblenden:
Schlagwörter:
-
Zusammenfassung:
Object recognition and categorization researchers have proposed that similarity may offer an organizational principle for the representation of objects in the brain and have demonstrated that such a principle is capable of supporting the distinct tasks of object recognition and categorization. But what is it that makes two objects perceptually similar? And what role does sensory modality play in shaping perceptual similarity? As a first step towards addressing this question, we used computer graphics modeling techniques to create novel, three-dimensional objects that varied parametrically in shape and texture.
Five levels of ‘texture blurring’ and five levels of ‘shape blurring’ were selected as the creation parameters for 25 objects. In a psychophysical experiment, fifteen human subjects viewed pairs of these objects and rated the similarity between them on a seven-point scale. Similarity ratings of the objects were also generated from a set of standard computer vision techniques based on pixel-wise differences, cross-correlations, edge images, and Gabor jets. Multidimensional scaling (MDS) was then performed on the similarity data. MDS on the perceptual similarities led to an optimal two-dimensional ‘perceptual map’ of the stimuli whose dimensions corresponded to texture and shape blur. Furthermore, the configuration of the stimuli
in this perceptual space closely resembled the configuration of the stimuli in the space defined by the creation parameters. Interestingly, the perceptual distance between stimuli along the shape dimension decreased as the amount of texturing increased, revealing an interaction
between shape and texture in the visual similarity judgments. The five texture levels were quite evenly spaced in the perceptual map, whereas the spacing between shape levels hinted at the formation of two distinct groupings based on shape. Comparisons against computational measures
revealed that simple pixel-based and correlation-based similarities were closest to the perceptual similarities, whereas similarities based on edge detection and Gabor jets overemphasized the importance of the texture dimension compared to perceptual similarities. This study establishes an experimental paradigm for generating information-rich maps of a set of stimuli and comparing maps derived from different similarity measures. Perceptual similarity
measures can be compared against computational similarity measures generated using various object features, providing an opportunity to evaluate the perceptual validity of those
features. In future work, this approach will be used to compare perceptual maps based on visual similarities against maps based on haptic similarities and in studies of multimodal categorization.
