Hourglass MobileNetV3_large Stereo Matching Network

Abstract

Stereo matching using rich contextual information can reduce false matches in pathological regions. Inadequate extraction of contextual feature information as well as high complexity matching models are the main reasons for the low accuracy and poor generalization of stereo matching models. To solve the above problems, we propose an Hourglass MobileNetV3_large stereo matching network. Firstly, the model is based on the lightweight MobileNetV3_large, which recovers and fuses feature information for high-resolution feature maps by combining feature maps at different scales. Secondly, the hybrid attention mechanism is embedded into the model to complete the further enhancement of the capture of global contextual information. The matching cost volume of lightweight group-wise correlation is utilized to complete the measurement of the difference between the corresponding pixels of the left and right images. Finally, the optimization of the parallax map is done by a modified hourglass-type regularization module. The model is tested and evaluated on SceneFlow, KITTI2012, KITTI2015, and Middlebury 2014. Experimental results show that the network model proposed in this paper even outperforms PSMNet algorithm and GwcNet algorithm on the KITTI2015 dataset in terms of All-D1-all metrics by 18.5% and 10.4%, respectively. Test results at Middlebury 2014 also show that the algorithm in this paper has good generalization performance. In addition, the network framework can also be applied to a variety of neural network models and good experimental results have been obtained for all of them.