Efficient Stereo Reconstruction with Uncertainty Estimation

The reconstruction of depth information from two stereoscopic images in an efficient way remains a dynamic field of research motivated by various applications, such as autonomous driving, robotics, and others. The uncertainty of the estimated depth can be valuable in practical applications and moreover, it can improve the estimated disparity in multi-task learning settings. This work proposes an enhancement of the 2D MobileStereoNet, including efficient uncertainty estimation as part of multitask learning that simultaneously addresses disparity and uncertainty. The proposed model is evaluated using a popular stereo images dataset, and comparisons with state of the art approaches illustrate its effectiveness.