Multi-scale inputs and context-aware aggregation network for stereo matching

Despite the significant progress made in deep learning-based stereo matching, the accuracy of these methods significantly decreases when faced with challenges such as occlusions, reflections, textureless areas, and scale variations. In this paper, we propose MSCANet, a novel stereo matching network that integrates multi-scale inputs and context-aware aggregation ability. MSCANet effectively integrates rich multi-scale feature information and exhibits context-aware capability, thereby enabling it to achieve superior performance. Firstly, a multi-scale aware fusion module is designed to efficiently incorporate more comprehensive global context features at different scales, which allows the model to enhance its ability to generalize across images of varying scales. Secondly, a novel V-shaped encoder/decoder module is developed to effectively exploit the rich feature information. In the encoding stage, a 3D squeeze-and-excitation block is introduced to facilitate adaptively recalibration of learned feature maps. This block effectively suppresses irrelevant features while enhancing useful features, which improved efficiency and accuracy in disparity prediction. Additionally, a 3D context-aware decode block is designed to effectively utilize global context features to restore the original image structure during the decoding stage. Moreover, the high-level feature maps can be employed to augment low-level feature maps by incorporating more detailed information to avoid the side effects caused by the loss of information during the encoding process. Extensive ablation experiments and comparative experiments were conducted on Scene Flow dataset, KITTI2012 and KITTI2015 datasets to validate the effectiveness of each proposed module. The experimental results demonstrate MSCANet achieves competitive performance and offers a more straightforward and efficient model design, as well as faster inference speed.