Multiscale Modeling and Constraints for Max-flow/Min-cut Problems in Computer Vision

Multiscale techniques have been used for many years in computer vision. Recently multiscale edges have received attention in spectral graph methods as an important perceptual cue. In this paper multiscale cues are used in the context of max-flow/min-cut energy minimization. We formulate multiscale min-cut versions of three typical computer vision applications, namely interactive segmentation, image restoration, and optical flow. We then solve across all scales simultaneously. This use of multiscale models and constraints leads to quantitatively and qualitatively improved experimental results.