Electronic Thesis and Dissertation Repository

Degree

Master of Science

Program

Computer Science

Supervisor

John Barron

Abstract

Optical flow is an important research area in the Computer Vision field, with the estimation of optical flow at occlusion still an open problem. Traditional approaches to this problem have either used additional terms in a regularization calculation (the flow still tends to “bleed” across occlusion boundaries) or a local least squares calculation that attempted to minimize the influence of two adjacent differently moving regions on the optical flow at points close to both regions (the flow still tends to be “corrupted” by the two regions). Ideally, optical flow for two adjacent differently moving regions should be distinct right up to the occlusion boundary.

A recent approach to calculate optical flow at occlusion is to combining boundary and region segmentation with the optical flow computation. Based on the work of Sundberg et al. Arbelaez et al. and Brox et al., we implement a motion gradient (mg) edge map algorithm which detects motion information in closed regions in the image sequences. Here we utilize the motion gradient as an additional local cue in the globalized probability of a boundary (gPb) as a new boundary detector to produce a gPb + mg contour map. The next step is to apply the Ultrametric Contour Map (UCM) mechanism, which is a framework to compute closed contours in a hierarchical region tree to produce a hierarchical edge map which indicates possible boundaries, including occlusion boundaries.

We implemented Sundberg et al.’s work to detect occlusion boundaries using optical flow, but, unlike Sundberg et al., we compute and display optical flow everywhere. The Sundberg et al. optical flow was generated by Brox et al’s method. They used a least squares calcula- tion on the brox flow at pixels around an occlusion boundary to determine whether a boundary computed by the gPb − UCM library developed by UC Berkeley is occluding or occluded. We extended their least squares idea to 1st and 2nd order optical flow models to generate dense opti- cal flow inside each closed region. Finally, we analyze our optical flow fields both qualitatively and quantitatively. In particular, for quantitative analysis, we use warping error, as the correct flow is unknown. We show improved results over those of Sundberg et al., note a number of shortcomings in Sundberg et al.’s approach and point to areas of future research.


Share

COinS