Learning Uncertainty For Safety-Oriented Semantic Segmentation In Autonomous Driving

In this paper, we show how uncertainty estimation can be leveraged to enable\nsafety critical image segmentation in autonomous driving, by triggering a\nfallback behavior if a target accuracy cannot be guaranteed. We introduce a new\nuncertainty measure based on disagreeing predictions as measured by a\ndissimilarity function. We propose to estimate this dissimilarity by training a\ndeep neural architecture in parallel to the task-specific network. It allows\nthis observer to be dedicated to the uncertainty estimation, and let the\ntask-specific network make predictions. We propose to use self-supervision to\ntrain the observer, which implies that our method does not require additional\ntraining data. We show experimentally that our proposed approach is much less\ncomputationally intensive at inference time than competing methods (e.g.\nMCDropout), while delivering better results on safety-oriented evaluation\nmetrics on the CamVid dataset, especially in the case of glare artifacts.\n