Learnable upsampling bilateral grid refinement for stereo matching network

Most existing binocular stereo matching algorithms require a trade-off between accuracy and speed, unable to achieve both simultaneously. One reason lies in the complexity and variability of scenes that stereo matching tasks must handle, where disparities in heavily textured, weakly textured, and occluded areas are often difficult to infer correctly. Therefore, this paper proposes the Learnable Upsampling Bilateral Grid Refinement for Stereo Matching Network (LUGNet). Through learnable bilateral grid upsampling guided by the left image, LUGNet calculates offsets for cost volume upsampling, while simultaneously leveraging the network to automatically learn interpolation weights to accommodate features of different datasets. Ultimately, LUGNet achieves error rates comparable to high-precision networks with a parameter count of 2.6M and an inference time of 58ms.