Gradient-Based Uncertainty for Monocular Depth Estimation

In monocular depth estimation, disturbances in the image context, like moving objects or reflecting materials, can easily lead to erroneous predictions. For that reason, uncertainty estimates for each pixel are necessary, particular safety-critical applications such as automated driving. We propose a post hoc estimation approach an already trained and thus fixed model, represented by deep neural network. The is estimated with gradients which extracted auxiliary loss function. To avoid relying on ground-truth information definition, we present function based correspondence of prediction its horizontally flipped counterpart. Our achieves state-of-the-art results KITTI NYU Depth V2 benchmarks without need retrain Models code publicly available at https://github.com/jhornauer/GrUMoDepth .