Philip Kellman¹, Austin Phillips¹, Nicholas Baker², Patrick Garrigan³; ¹University of California, Los Angeles, ²Loyola University Chicago, ³St. Joseph's University

How does visual processing transform early encodings—transient responses of orientation-sensitive units—into durable, symbolic representations of object boundaries and shape? Recent research provides clues about how this transition occurs in contour and two-dimensional shape perception. Evidence suggests that an initial symbolic representation of contour shape may consist of segments of constant curvature (Baker, Garrigan, & Kellman, 2021). We hypothesize that these representations arise from banks of constant-curvature filters ("arclets") built from oriented units linked by constant turning angles (Kellman, et al., 2013; c.f., Poirier & Wilson, 2006). Arclets span multiple turning angles and scales; collectively, they force any smooth contours into a representation of constant curvature segments. Baker & Kellman (2021) developed a computational model supporting the plausibility of the general approach and its agreement with psychophysical data. Here, we describe a neurally plausible model that operates on simple images and produces a symbolic encoding of open contours and 2D shapes in terms of constant curvature segments. Each arclet is composed of three co-circular, odd-symmetric, Gabor filters. We use 6 scales, with 6 positive and 6 negative turning angles, plus a zero turning angle, per scale. Filled 2D shapes on homogeneous backgrounds are convolved with each filter type. Activations along the contour, normalized in terms of maximum possible activation, are measured for the best fitting arclet of each type. Arclets of the various turning angle and scale combinations compete to "capture" local contour segments. Model tests show that the resulting representations agree well with human perception, even for arbitrary shapes lacking constant-curvature parts. We further describe how encoding with curvature filters yields a scale-invariant shape code that may underlie the ease with which perceivers detect shape identity and similarity across in absolute size.

Acknowledgements: NIH R01CA236791 to PK