Towards Pixel-Level OOD Detection for Semantic Segmentation

There exists wide research surrounding the detection of out of distribution sample for image classification. Safety critical applications, such as autonomous driving, would benefit from the ability to localise the unusual objects causing an image to be out of distribution. This thesis adapts state-of-the-art methods for detecting out of distribution images for image classification to the new task of detecting out of distribution pixels, which can localise the unusual objects. It further experimentally compares the adapted methods to a new dataset derived from existing semantic segmentation datasets, proposing a new metric for the task. The evaluation shows that the performance ranking of the compared methods successfully transfers to the new task.