A selection model for motion transparency in visual cortex

Abstract

Purpose: Identify cortical mechanisms for the perception of motion transparency and motion occlusion using a computational model based on the response properties of neurons in visual cortex. Methods: The inputs to the model are 64 frames of moving stimuli on a 64 x 64 pixel input array with 256 gray levels at each pixel. Area V1 was modeled by a 49x49 aray of 36 Adelson-Bergen motion energy units, tuned to 4 directions and 9 combinations of spatial and temporal frequencies. The model of MT had two parallel pathways, each receiving inputs from the motion-energy array. The first pathway was optimized to estimate the local velocity while the second pathway was optimized to select the most reliable regions to support each global velocity. The output was 33 global velocity units each receiving inputs from local velocity units gated multiplicatively by the selection units. Results: We tested the model on a variety of psychophysical motion stimuli including plaids, random dots, and barber-pole stimuli, and compared the response properties of the units in the model to those found in area MT. Plaid patterns cohered if the spatial requency of the component gratings were within one octave. When the contrast of the intersection was varied, the range of contrasts that yielded component motion corresponded to the range over which humans perceive motion transparency (Stoner, Ramachandran, and Albright, 1990). The threshold for detecting coherent motion in correlated random dots was similar to that of primates. Responses to the barber-pole illusion and rotation elipses were similar to those reported by humans. Conclusions: We suggest that a subpopulation of neurons in area MT is primarily concerned with robust selection rather than representing velocity. In visual scenes with occluded and transparent objects, the selection units partially segment the image into disjoint regions sharing a common global velocity. Attention to a particular velocity in a particular region of the visual field could be accomplished through bias of the selection units. The selection units had opponent nonclassical receptive fields, similar to those observed in some area MT neurons.