In the real-world deployment, computer vision models, such as object detectors and autonomous vehicles, always encounter data that are unlike the training data. Such data, also known as out-of-distribution data, can cause great harm because neural networks typically cannot make reliable predictions on them. The aim of this PhD project is to explore ways to help computer vision models to be able to deal with such data in terms of recognizing them and making informed predictions on them. Specifically, we will explore various methods like deep generative models, incorporating geometric cues, and instilling knowledge about the 3D world. Our final goal is to develop a systematic, effective, and efficient framework for computer vision models to generalize to out-of-distribution data.