Natural Visual Input Implies Estimation Biases in a Blowfly Motion Sensitive Neuron

Visual estimation of self-motion has been shown to be computationally similar across a wide variety of visual morphologies. Surprisingly, one of the conserved similarities is strong estimator biases in low signal-to-noise ratio regimes. Given the behavioral importance of motion estimation, the widespread existence of this bias suggests it arose as a feature of the statistics of natural scenes. Previously, we computationally investigated this phenomenon by using a specially designed "FlEye" camera made to mimic the fly visual system. We demonstrated that an optimal motion estimator constructed using scenes captured by the FlEye camera has naturally appearing correlator-like motion estimation properties at low signal-to-noise ratios. Additionally, we demonstrated that pitch acts as a source of noise during yaw estimation, effectively extending the range where correlator-like motion estimation is observed. These results suggest that the presence of "visual noise" drives biological motion estimation away from gradient-like and towards correlator-like estimators. Here, we connect these results back to the blowfly (Calliphora vicina) by comparing our computational motion estimator to H1, a wide field motion sensitive neuron involved in motion control in the fly: To do this, we performed playback experiments with FlEye videos using a high-intensity LED display while recording from H1. By comparing the behavior of the H1 neuron to our computational estimator, we characterize which features of the statistics of natural scenes are important for driving H1 toward correlator-like behaviors. Preliminary evidence shows that the biological motion sensor exhibits estimator biases that are qualitatively consistent with those predicted from measurements with the camera.