Guided aggregation and disparity refinement for real-time stereo matching

Stereo matching methods based on convolution neural network (CNN) often face challenges such as edge blurring and the loss of small structures. These issues often result in incorrect disparity assignments when upsampling the disparity map. To address this problem, we propose a disparity refinement module (GDU-CTF) that combines guided disparity map upsampling with a coarse-to-fine process. This approach effectively restores incorrect disparity values in the final disparity map. Furthermore, due to the insufficient aggregation of global geometric and contextual texture features using basic encoder–decoder 3D convolutional networks, we propose a guided patch cost aggregation module (GPA) that generates a more precise initial disparity map for textureless areas. These modules complement each other and are efficient, resulting in an accurate and lightweight framework for stereo matching. Experimental results demonstrate that our algorithm has excellent accuracy in generating disparity maps and achieves outstanding real-time performance, with an inference time of just 0.03 s on Scene Flow and KITTI datasets.