Linking model predictions to specific locations in space and time, such as object masks in particular video frames or temporal segments.