Dense Optic Flow with a Bayesian Occlusion Model

This paper presents a dense optic flow algorithm based on quite simple local probability assumptions. Due to the explicit derivation of the correspondence concept in a probability-theoretical framework, occlusion probability evolves straight-forwardly from the model for each pixel. Initialized with a similarity measure based on single pixels, an iterated diffusion step propagates local information across the image, while occlusion probability is used to inhibit flow information transfer across depth discontinuities, which prevents flow smoothing at 3d object boundaries. The inhibition is thereby not artificially modelled by some heuristically chosen parameters, but arises directly from the Bayesian correspondence model. The algorithm structure can be interpreted as a recurrent neural network, where matched points have reached a stable state, while others (e.g. those in homogeneous areas) keep receiving information from regions more and more far away until they converge, this way overcoming the aperture problem. The massive parallel structure allows for and demands a real hardware implementation of the system.