AGO: Accelerating Global Optimization for Accurate Stereo Matching

In stereo matching, global algorithms could produce more accurate disparity estimation than aggregated ones. Unfortunately, they remain facing prohibitively high computational challenges while minimizing an energy function. Although various computationally tractable optimizers such as Belief Propagation (BP), Graph Cut (GC) and Dynamic Programming (DP) exist, they still time-consumed and perform a relative higher energy. On one hand, too many intermediate parameters are required for constructing the energy function. In fact, they are trivial, difficult to compute and inevitably bring noises because of their approximate representations. That signifies a simplified energy function structure with fewer intermediate parameters may not only accelerate the running time but also provide greater result while utilizing same optimizer. On the other hand, optimizers are usually designed artificially and also generate approximate solutions. Based on this observation, a suboptimal energy is probably obtained. To alleviate these limitations, an Accelerating Global Optimization (AGO) stereo matching algorithm is proposed in this paper. Integrating the key ideas of Two-Step Global Optimization (TSGO), the modeled energy function of AGO possesses fewer intermediate parameters. What's more, a refinement term is augmented into the message passing formula of Sequential Tree-Reweighted (TRW-S) optimizer to lower down the energy. Performance evaluations on Middlebury v.2 & v.3 stereo data sets demonstrate that the proposed AGO outperforms than other five challenging stereo algorithms; and also performs better on Microsoft i2i stereo videos. At last, thanks to the simple energy function model of AGO; it shows a faster execution time.