Jingzhi Li

Master Thesis
Supervisors: Dr. Radu Timofte

Optical Flow with Improved Dense Inverse Search

In the constant search for a fast and accurate estimation technique for optical flow, the majority of work in the literature engage in the pursuit for high accuracy. Meanwhile, the fast-growing need for real-time optical flow application demands for high speed processing. We investigate the fast Dense Inverse Search (DIS) approach and aim to improve its accuracy without sacrificing speed. Four novel modifications are proposed: the bipolar search, the variance weighted patch- averaging scheme, the variance guided bilateral filtering and the selective dropout in variational refinement. A combination of the modifications are selected based on the application scenario. For small displacements, the proposed solution is able to run at comparable speed with a 16% increase in accuracy on the fastest setting.