Multi-Scale Generalized Plane Match for Optical Flow

TL;DR

This paper introduces a multi-scale generalized plane matching framework for optical flow estimation that improves accuracy and occlusion localization, especially in challenging scenarios with illumination changes and occlusions.

Contribution

It presents a novel multi-scale plane matching approach that models scenes as multiple planes, enhancing dense correspondence and occlusion detection in optical flow estimation.

Findings

Robust optical flow estimation on MPI-Sintel datasets

Accurate occlusion localization and thin object detection

Comparable flow quality with improved occlusion handling

Abstract

Despite recent advances, estimating optical flow remains a challenging problem in the presence of illumination change, large occlusions or fast movement. In this paper, we propose a novel optical flow estimation framework which can provide accurate dense correspondence and occlusion localization through a multi-scale generalized plane matching approach. In our method, we regard the scene as a collection of planes at multiple scales, and for each such plane, compensate motion in consensus to improve match quality. We estimate the square patch plane distortion using a robust plane model detection method and iteratively apply a plane matching scheme within a multi-scale framework. During the flow estimation process, our enhanced plane matching method also clearly localizes the occluded regions. In experiments on MPI-Sintel datasets, our method robustly estimated optical flow from given noisy correspondences, and also revealed the occluded regions accurately. Compared to other state-of-the-art optical flow methods, our method shows accurate occlusion localization, comparable optical flow quality, and better thin object detection.