Natural motion statistics unify direction computations

Abstract

Visual direction selectivity is typically studied using coherent, unidirectional stimuli. However, natural vision involves more complex, heterogeneous motion fields generated by self-motion and moving objects. How visual circuits extract direction information from these motion fields is not fully understood. Here, we show that On-Off direction-selective ganglion cells (DSGCs) in the mouse retina encode an integrated motion direction from component motion vectors within their receptive fields. Using plaid stimuli with distinct velocity components, we demonstrate that DSGCs maintain unidirectional tuning even when component velocity vectors diverge by over 90 degrees. DSGCs preferred directions follow the angular mean, rather than the vector sum of motion vectors, a property emerging from the directional tuning of their inhibitory inputs. Intriguingly, while angular mean and vector sum represent distinct mathematical operations, the statistical properties of natural movies cause them to converge onto a single aligned direction, eliminating the necessity of the visual system to choose one operation over the other. We found that DSGCs encode this aligned direction for both optic flow and object motion movies, and the precision of direction encoding depends on the homogeneity of motion vectors in the receptive field. Our findings reveal that the retina performs an initial integration of motion vectors using angular averaging. Since natural motion statistics cause possible motion vector computations to converge, angular averaging effectively informs the animal about the consensus motion direction in natural environments.