Machine learning has enabled major advances in the detection and tracking of objects that are visible to the camera. However, some use cases for autonomous systems (e.g. self-driving cars) require the ability to reason about occluded objects that disappear from view – before they reappear. Through their proposed benchmark and metrics, Khurana et al. find that most detection and tracking systems cannot adequately track people (e.g. pedestrians) through occlusion. They propose extensions to linear dynamic models to forecast object trajectories during occlusion and leverage monocular depth estimation (to infer depth from 2D images) to address this limitation.