Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing data characterizing a first region in an environment to generate a prediction characterizing a second region in the environment. One of the methods includes obtaining surfel data comprising a plurality of surfels; obtaining sensor data for a plurality of locations in a first region of the environment; determining, from the surfel data, a plurality of first surfels corresponding to respective locations in the first region of the environment; determining, using the first surfels and the sensor data, a difference between i) a first representation of the first region of the environment corresponding to the first surfels and ii) a second representation of the first region of the environment corresponding to the sensor data; and generating a respective object prediction for one or more locations in a second region of the environment.