In one embodiment, a computing system accesses a number of features extracted from one or more first images. The extracted features are associated with at least one object captured in the first images. The first images are captured by a camera associated with a vehicle. The computing system identifies, in a reference map, reference features matching one or more of the features extracted from the first images. The reference features are associated with the at least one object captured in the first images. The computing system generates, for the camera, a calibration model by comparing the identified reference features in the reference map and the features that match the one or more reference features. The calibration model is used to calibrate second images captured by the camera associated with the vehicle.