MCVFNet: Multi-Information Cost Volume Fusion for Real-Time Stereo Matching

Stereo matching-based depth estimation that provides dense 3D information is crucial to perceive the surrounding environment. The majority of current stereo matching networks aim to improve prediction accuracy by leveraging robust regularization through extensive use of 3D convolutions, but at the cost of speed. Thus, we propose MCVFNet (Multi-Information Cost Volume Fusion) to reduce the need for 3D convolution through the addition of multiple information. We propose BAF (Bidirectional Aggression Fusion) and MAF (Multi-scale Attention Fusion) for multi-information fusion in the cost aggregation. BAF effectively incorporates multi-scale information into final cost volume without need for excessive 3D convolutions. Through the interconnection of the above and below information streams, which acquires more contextual and geometric information, meanwhile, we propose MAF to fully integrate the information contained in other multi-size cost volumes into the final cost volume. Effective refinement of the cost volume aggregated by BAF is achieved in both disparity dimension and spatial dimension. Experimental results show that we achieve real-time performance on the NVidia 2080TI, outperforming most other recent methods on KITTI online validation set, and competitive with state-of-the-art models.