MStereoNet: A Lightweight Stereo Matching Network Using MobileNet

Abstract

Deep learning model-based approaches to stereo matching challenges are more accurate than conventional feature-based techniques created by hand. This leads to the issue that deploying applications on devices with restricted resources is not friendly to employing complicated networks and total cost space to increase performance. To minimize processing effort without sacrificing matching accuracy, we propose MStereoNet in this study, a more effective stereo network. It has been demonstrated experimentally that the network in this research significantly lowers the requirement for computing power. The network in this article also employs a multiscale loss to enhance the reliability of the detail. The approach in this study delivers a performance comparable to the best existing algorithms in the Sceneflow dataset when compared to other low-cost dense stereo depth estimation techniques. Research demonstrates that the network suggested in this study can reduce up to 72.5% and 87% of the parameters and operations than the largest volume of the methods involved in the comparison. Our network also marginally outperforms other lightweight binocular matching networks in terms of accuracy.