MAFNet: Multi-scale Attention Fusion Network for Enhanced Stereo Matching

Stereo matching, crucial for applications in autonomous driving and 3D reconstruction, faces challenges due to model redundancy and limited adaptability to complex scenes. This paper introduces MAFNet, a novel stereo matching network that integrates multi-scale cost volumes and attention mechanisms. MAFNet utilizes a multi-scale cost stepped fusion method to combine cost volumes of different resolutions, effectively expanding the model's receptive field. An attention weight generator is employed to filter redundant features, enhancing the conciseness and effectiveness of the cost volume. Furthermore, an information fusion hourglass network combines contextual and spatial information, leading to improved depth estimation accuracy. On the KITTI 2015 dataset, MAFNet achieves a percentage of stereo disparity outliers as low as 1.76%, ranks among the best in similar binocular stereo matching networks that integrate multi-scale information fusion.