Abstract
In this paper, we tackle the detection of out-of-distribution (OOD) objectsin semantic segmentation. By analyzing the literature, we found that currentmethods are either accurate or fast but not both which limits their usabilityin real world applications. To get the best of both aspects, we propose tomitigate the common shortcomings by following four design principles:decoupling the OOD detection from the segmentation task, observing the entiresegmentation network instead of just its output, generating training data forthe OOD detector by leveraging blind spots in the segmentation network andfocusing the generated data on localized regions in the image to simulate OODobjects. Our main contribution is a new OOD detection architecture calledObsNet associated with a dedicated training scheme based on Local AdversarialAttacks (LAA). We validate the soundness of our approach across numerousablation studies. We also show it obtains top performances both in speed andaccuracy when compared to ten recent methods of the literature on threedifferent datasets.