Devon: Deformable Volume Network for Learning Optical Flow.

We propose a lightweight neural network model, Deformable Volume Network(Devon) for learning optical flow. Devon benefits from a multi-stage frameworkto iteratively refine its prediction. Each stage is by itself a neural networkwith an identical architecture. The optical flow between two stages ispropagated with a newly proposed module, the deformable cost volume. Thedeformable cost volume does not distort the original images or their featuremaps and therefore avoids the artifacts associated with warping, a commondrawback in previous models. Devon only has one million parameters. Experimentsshow that Devon achieves comparable results to previous neural network models,despite of its small size.