Analysis of 3D reconstruction error in the context of computational stereo in remote sensing

The availability of stereo couples of satellite or aerial images allows for the extraction of the Digital Surface Model which is of high importance for earth observation. In the specific case of change detection of man made structures, the 3D information is a vital cue to distinguish buildings from flat objects (e.g. parking lots). Height variations are obtained from a disparity map, qualifying the magnitude of the parallax and henceforth the relative elevation. If available, the camera and flight parameters enable other types of applications thanks to the derivation of geographical 3-D coordinates. In particular, the National Geographic Institute of Belgium is interested in automatic estimation of 3D coordinates to support the nowadays time consuming manual procedure. In a first collaboration with NGI, we developed two disparity estimation procedures to detect buildings and look for related changes in the topographic database. Then flight and camera data were used to derive geographical 3D coordinates. This paper first recalls how the different steps necessary to produce a DSM from stereo images were implemented, addressing epipolar rectification, disparity estimation and 3D reconstruction. The major contribution of the paper is to present a quantitative assessment of the quality of automatic 3D measures by comparison with available ground truth values. A separation between errors from artefacts in disparity estimation and precision limits in the reconstruction process has been considered.