Benchmarking photon-limited performance of optic flow processing algorithms

In this paper we present a simulation framework for testing and benchmarking the photon-limited performance of optic flow processing techniques. We explore the performance of “traditional” gradient-based and feature-based optic flow algorithms as well as the “biologically-inspired” elementary motion detector. We show that biologically-inspired spatial pooling techniques can successfully be implemented in the gradient-based image interpolation algorithm and the elementary motion detector to extend low-light capabilities by more than one order of magnitude. We also show that block matching algorithms can accurately compute optic flow even when a single frame may capture <;1000 photons. This framework provides a tool to further explore the mechanisms that underlie the observed low-light performance of biological vision systems and to understand the relative performance of different approaches for implementing optic flow in hardware.