Deep learning based stereo matching algorithms have produced impressive disparity estimation for recent years; and the success of them has once overshadowed the conventional ones. In this paper, we intend to reverse this inferiority, by leveraging Stacking Learning with Coalesced Cost Filtering to make the conventional algorithms achieve or even surpass the results of deep learning ones. Four classical and Discriminative Dictionary Learning (DDL) algorithms are adopted as base-models for Stacking. For the former ones, four classical stereo matching algorithms are employed and regarded as ‘Coalesced Cost Filtering Module’; for the latter supervised learning one, we utilize the Discriminative Dictionary Learning (DDL) stereo matching algorithm. Then three categories of features are extracted from the predictions of base-models to train the meta-model. For the meta-model (final classifier) of Stacking, the Random Forest (RF) classifier is selected. In addition, we also employ an advanced one-view disparity refinement strategy to compute the final refined results more efficiently. Performance evaluations on Middlebury v.2 and v.3 stereo data sets demonstrate that the proposed algorithm outperforms other four most challenging stereo matching algorithms. Besides, the submitted online results even show better results than deep learning ones.

近年来，基于深度学习的立体匹配算法产生了令人印象深刻的视差估计；他们的成功一度让传统的黯然失色。在本文中，我们打算通过利用 Stacking Learning 和 Coalesced Cost Filtering 来扭转这种劣势，使传统算法达到甚至超过深度学习算法的结果。采用四种经典的判别字典学习 (DDL) 算法作为 Stacking 的基础模型。对于前者，采用四种经典的立体匹配算法，并将其视为“合并成本过滤模块”；对于后一种监督学习，我们利用判别字典学习 (DDL) 立体匹配算法。然后三个从基础模型的预测中提取特征类别来训练元模型。Stacking的元模型（最终分类器）选择随机森林（RF）分类器。此外，我们还采用了一种先进的单视图视差细化策略来更有效地计算最终的细化结果。Middlebury v.2和v.3立体数据集的性能评估表明，所提出的算法优于其他四种最具挑战性的立体匹配算法。此外，提交的在线结果甚至显示出比深度学习更好的结果。