Abstract
Introduction. Curvature encoding serves as an intermediate step towards the neural representation of shapes and object parts (Loffler, Wilson & Wilkinson, 2003, Vision Research; Wilkinson, Wilson & Habak, 1998, Vision Research). We compared the performance of several curvature-encoding schemes (inspired by computational, physiological, and psychophysical considerations) with respect to natural constraints imposed by shape perception tasks.
Methods. Using an image filtering approach, we evaluate the response properties of different curvature encoding schemes, with respect to (1) object size variability for size constancy, (2) curvature amplitude, (3) response noise away from regions of stimulus curvature, and (4) 1st order to 2nd order contour alignment for application to texture edges or 2nd order contours.
Results. Results indicate that: (1) to combine successive edge elements, an “AND” operator is preferable to a linear sum filter to reduce neural noise away from loci of maximum curvature, (2) filter properties need to be adjusted to object size otherwise systematic distortions in the locations of response maxima may occur, (3) introducing orientation-selectivity to the inputs to the curvature mechanism only modestly sharpens curvature responses in well-defined isolated contours, and (4) opponent-curvature mechanisms have greater spatial- and curvature-selectivity.
Discussion. By developing an understanding of success and failures of different mechanisms, we isolated the requirements of neural curvature mechanisms. We provide important constraints on the design of biologically plausible curvature filters for use in object processing models. We also discuss a fast method of implementing position-dependent curvature mechanisms that can be used to recover curvature responses independent of object size.
NSERC grant to HRW (#OP227224)