Abstract
The visual representation of shape condenses a high-dimensional input image into a smaller set of more informative features. One of the main functions of such representations is to support similarity judgments, so that different shapes can be related to one another and significant features identified. Although distinguishing the causal origin of features is under-constrained, ideally, shapes resulting from similar underlying generative process (e.g. subjected to similar transformations) should be seen as more similar to one another than those created by different processes. To test this, we created sets of novel 2D objects, based on a small number of naturalistic base shapes. Half the shapes in each set were morphs of the base shapes. These non-transformed 'source' shapes were constant across sets. The other half of each set was created by subjecting the source objects to shape-deforming transformations (stretching or bloating to different degrees). Using a multi-arrangements method, observers spatially arranged each set of shapes according to their similarity within a circular arena. The results indicate that transformed shapes were seen as a distinct category from the source objects, but were also arranged in an orderly fashion, closest to their non-transformed counterparts. Thus, participants arranged the sets according to both the source and the transformation, suggesting that they could segment shape features according to their causal origins. Moreover, the larger transformations caused shapes to converge in shape space, suggesting that some generative processes create distinctive signatures in shapes independent of their original form. Representational similarity analysis with mid-level shape features (e.g., contour/skeletal structure) showed that similarities in the transformation and contour of the shapes best account for the participants' arrangements. These results suggest that the visual system represents shapes using context-dependent feature space that takes into account the causal origin of different features.
Meeting abstract presented at VSS 2017