This paper describes a system to generate 3D model from image sequence taken in complex environment including variable ground surface, buildings and trajectory loops. Here we use a $1000 catadioptric camera and an approximate knowledge of its calibration. This contrasts to current systems which rely on more costly hardware such as the (calibrated) spherical vision camera Ladybug. All steps of the method are summarized. Experiments include a campus reconstruction from thousands of images.