State-of-the-art deep neural networks (DNNs) are usually trained to operate on a pre-defined and closed set of categories. Thus, they are ill-equipped when exposed to examples of a novel and unknown category. In real-world applications, to which DNNs are envisioned to be deployed to, the missing capability of handling such so-called “out-of-distribution” (OoD) examples could potentially lead to unwanted consequences. This becomes particularly crucial in high-stake applications. In this project, we tackle the problem of OoD detection in semantic segmentation, which is a key perception component in many existing vision systems based on machine learning. To this end, we are going to employ generative models to evaluate the likelihood of latent representations extracted from internal layers of the encoder of semantic segmentation DNNs. The intuition is that the likelihood measures how well observed features fit to those already known from training. We investigate and develop multiple types of generative models, paying special attention to the challenge of handling very high dimensional input data.