Dual Cost Volume Stereo Matching Method Based on Gated Cost Aggregation

Existing stereo matching algorithms often construct a single large cost volume, requiring stacked deep convolutional networks during the cost aggregation stage, which incurs high computational costs. This paper proposes a novel stereo matching algorithm that builds two small cost volumes instead of a single global one, computing only in key regions or scales, with the two cost volumes undergoing independent cost aggregation. During the cost aggregation stage, a gated spatial convolution module with skip connections is introduced, which effectively filters irrelevant information while preserving or amplifying valid matching cues. A fusion module is proposed to interactively integrate matching information extracted from the two distinct cost volumes. The network achieves strong accuracy and speed on the SceneFlow, KITTI2012, and KITTI2015 datasets, with an EPE of 0.68 on SceneFlow dataset and 0.41 on the mixed KITTI2012 and KITTI2015 datasets. Disparity map inference time on the KITTI dataset is 100 milliseconds.