We argue that most natural objects have a part structure that we can recover from image data and thus use as the basis for general-purpose recognition. We describe a parts representation that is fairly general purpose, despite having only a small number of parameters. Having this expressive power captured by a small number of parameters allows us to approach the problem of recovering an object's part structure by use of minimal length encoding. We present several examples of recovering part structure using various types of range imagery to show that the recovery procedure is robust.