Uncertainty estimation for stereo matching based on evidential deep learning

Although deep learning-based stereo matching approaches have achieved excellent performance in recent years, it is still a non-trivial task to estimate the uncertainty of produced disparity map. In this paper, we propose novel approach both aleatoric and epistemic uncertainties for an end-to-end way. We introduce evidential distribution, named Normal Inverse-Gamma (NIG) whose parameters can be used calculate uncertainty. Instead directly regressed from aggregated features, are predicted each potential then averaged via guidance probability distribution. Furthermore, considering sparsity ground truth real scene datasets, design two additional losses. The first one tries enlarge on incorrect predictions, so becomes more sensitive erroneous regions. second enforces smoothness regions with smooth disparity. Most models, such as PSM-Net, GA-Net, AA-Net, easily integrated our approach. Experiments multiple benchmark datasets show that method improves results. prove well-calibrated predictions. Particularly, capture increased out-of-distribution data, making effective prevent system fatal consequences. Code available at https://github.com/Dawnstar8411/StereoMatching-Uncertainty.