Identifying Out-of-Domain Objects with Dirichlet Deep Neural Networks

Ahmed Hammam, Frank Bonarens, Seyed Eghbal Ghobadi, Christoph Stiller; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2023, pp. 4560-4569

Deep neural networks are usually trained on a closed set of classes, which makes them distrustful when handling previously-unseen out-of-domain (OOD) objects. In safety-critical applications such as perception for automated driving, detecting and localizing OOD objects is crucial, especially if they are positioned in the driving path. In the context of this contribution, OOD objects refer to objects that were not represented in the training dataset. We propose a Dirichlet deep neural network for instance segmentation with inherent uncertainty modeling based on Dirichlet distributions and the Intermediate Layer Variational Inference (ILVI). A thorough analysis shows that our method delivers reliable uncertainty estimates to its predictions whilst identifying OOD instances. The model-agnostic approach can be applied to different instance segmentation models as demonstrated for two different state-of-the-art deep neural networks. Superior results can be shown on the out-of-domain Lost and Found dataset compared to state-of-the-art approaches, whilst also achieving improvements on the in-domain Cityscapes dataset.