Displacement-Invariant Cost Computation for Stereo Matching

Abstract Although deep learning-based methods have dominated stereo matching leaderboards by yielding unprecedented disparity accuracy, their inference time is typically slow, i.e. , less than 4 FPS for a pair of 540p images. The main reason that the leading employ time-consuming 3D convolutions applied to 4D feature volume. A common way speed up computation downsample volume, but this loses high-frequency details. To overcome these challenges, we propose displacement-invariant cost module compute costs without needing Rather, are computed applying same 2D convolution network on each disparity-shifted map independently. Unlike previous convolution-based simply perform context mapping between inputs and maps, our proposed approach learns match features two We also an entropy-based refinement strategy refine map, which further improves avoiding need second right image. Extensive experiments standard datasets (SceneFlow, KITTI, ETH3D, Middlebury) demonstrate method achieves competitive accuracy with much time. On typical image sizes ( e.g. $$540\times 960$$ 540 × 960 ), processes over 100 desktop GPU, making suitable time-critical applications such as autonomous driving. show generalizes well unseen datasets, outperforming 4D-volumetric methods. will release source code ensure reproducibility.