Fool Me Once: Robust Selective Segmentation via Out-of-Distribution Detection with Contrastive Learning

In this work, we train a network to simultaneously perform segmentation and\npixel-wise Out-of-Distribution (OoD) detection, such that the segmentation of\nunknown regions of scenes can be rejected. This is made possible by leveraging\nan OoD dataset with a novel contrastive objective and data augmentation scheme.\nBy combining data including unknown classes in the training data, a more robust\nfeature representation can be learned with known classes represented distinctly\nfrom those unknown. When presented with unknown classes or conditions, many\ncurrent approaches for segmentation frequently exhibit high confidence in their\ninaccurate segmentations and cannot be trusted in many operational\nenvironments. We validate our system on a real-world dataset of unusual driving\nscenes, and show that by selectively segmenting scenes based on what is\npredicted as OoD, we can increase the segmentation accuracy by an IoU of 0.2\nwith respect to alternative techniques.\n