Fast and accurate multiscale feature fusion stereo matching network

Abstract

In this paper, we propose a novel deep learning network model for stereo matching that achieves higher accuracy and better real-time performance. We achieve the effect of streamlining the complexity of the network model by using the lightweight network model MobileNetV3 for feature extraction. By using feature maps at four different scales, combined with the attention feature blocks in the network, the network can combine contextual information with better sensing ability, and the network model achieves a more accurate and efficient processing effect through multiscale feature fusion. We pretrained the model on the SceneFlow virtual datasets and fine-tuned it with multiple epochs of training on the KITTI datasets, which resulted in better generalization of the stereo matching model and more accurate disparity prediction results for real traffic and road environments. Compared with some existing network models such as PSMNet, GANet, our network achieves a faster processing rate with guaranteed accuracy, which meets the real-time requirements of network models under general conditions.