Volumetric Correspondence Networks for Optical Flow

Our innovations dramatically improve accuracy over SOTA on standard flow benchmarks while being significantly easier to work with - training converges in 7X fewer iterations. Interestingly, our networks appear to generalize across diverse correspondence tasks. On-the-fly adaptation of search windows allows ustorepurpose optical flownetworks for stereo (and vice versa), and can also beused toimplement adapativenetworks that increase search windowsizeson-demand.