Uncertainty Estimation for End-To-End Learned Dense Stereo Matching via Probabilistic Deep Learning

Motivated by the need to identify erroneous disparity assignments, various\napproaches for uncertainty and confidence estimation of dense stereo matching\nhave been presented in recent years. As in many other fields, especially deep\nlearning based methods have shown convincing results. However, most of these\nmethods only model the uncertainty contained in the data, while ignoring the\nuncertainty of the employed dense stereo matching procedure. Additionally\nmodelling the latter, however, is particularly beneficial if the domain of the\ntraining data varies from that of the data to be processed. For this purpose,\nin the present work the idea of probabilistic deep learning is applied to the\ntask of dense stereo matching for the first time. Based on the well-known and\ncommonly employed GC-Net architecture, a novel probabilistic neural network is\npresented, for the task of joint depth and uncertainty estimation from epipolar\nrectified stereo image pairs. Instead of learning the network parameters\ndirectly, the proposed probabilistic neural network learns a probability\ndistribution from which parameters are sampled for every prediction. The\nvariations between multiple such predictions on the same image pair allow to\napproximate the model uncertainty. The quality of the estimated depth and\nuncertainty information is assessed in an extensive evaluation on three\ndifferent datasets.\n