Massively Parallel Optical Flow using Distributed Local Search

摘要

The design of many tasks in computer vision field requires addressing difficult NP-hard energy optimization problems. An example of application is the visual correspondence problem of optical flow, which can be formulated as an elastic pattern matching optimization problem. Pixels of a first image have to be matched to pixels in a second image while preserving elastic smoothness constraint on the first image deformation. In this paper, we present a parallel approach to address optical flow problem following the concept of distributed local search. Distributed local search consists in the parallel execution of many standard local search processes operating on a partition of the data. Each process performs local search on its own part of the data such that the overall energy is minimized. The approach is implemented on graphics processing unit (GPU) platform and evaluated on standard Middlebury benchmarks to gauge the substantial acceleration factors that can be achieved in the task of energy minimization.