Abstract
The perception of object motion requires the integration of local estimates of image motion across space. The two general computational strategies that have been offered to explain spatial integration can be classified as hierarchical or lateral interactive. The hierarchical model assumes local motion estimates at a lower point in the hierarchy are integrated by neurons with large receptive fields. These neurons could make use of the fact that due to the aperture problem the 2D distribution of local velocities for a rigid translation falls on a circle through the origin in velocity space. However the challenge for this approach is how to segment and represent the motion of different objects or textures falling within the receptive field, including how to represent object boundaries. Apparent global rotations and dilations can be instantiated in randomly oriented global Gabor arrays suggesting that the aperture problem can be resolved though local interactions. The challenge for this approach is to discover local rules that will allow global organizations to emerge. These rules need to incorporate the status of ambiguous motion signals and unambiguous motion signals to explain how unambiguous 2D motion cues (e.g. at corners) influence the computed global motion field. Here we will describe a simple least squares approach to local integration, demonstrate its effectiveness in dealing with the dual problems of integration and segmentation and consider its limitations.