Stereo image dense matching, which plays a key role in 3D reconstruction, remains a challenging task in photogrammetry and computer vision. In addition to block-based matching, recent studies based on artificial neural networks have achieved great progress in stereo matching by using deep convolutional networks. This study proposes a novel network called a dual guided aggregation network (Dual-GANet), which utilizes both left-to-right and right-to-left image matchings in network design and training to reduce the possibility of pixel mismatch. Flipped training with a cost volume consistentization is introduced to realize the learning of invisible-to-visible pixel matching and left−right consistency matching. In addition, suppressed multi-regression is proposed, which suppresses unrelated information before regression and selects multiple peaks from a disparity probability distribution. The proposed dual network with the left−right consistent matching scheme can be applied to most stereo matching models. To estimate the performance, GANet, which is designed based on semi-global matching, was selected as the backbone with extensions and modifications on guided aggregation, disparity regression, and loss function. Experimental results on the SceneFlow and KITTI2015 datasets demonstrate the superiority of the Dual-GANet compared to related models in terms of average end-point-error (EPE) and pixel error rate (ER). The Dual-GANet with an average EPE performance = 0.418 and ER (>1 pixel) = 5.81% for SceneFlow and average EPE = 0.589 and ER (>3 pixels) = 1.76% for KITTI2005 is better than the backbone model with the average EPE performance of = 0.440 and ER (>1 pixel) = 6.56% for SceneFlow and average EPE = 0.790 and ER (>3 pixels) = 2.32% for KITTI2005.
Keywords: deep learning; dense image matching; left–right consistency.