Jamet - 6 
whether an adaptation of this technique could lead to cartographic contours, or to the 
equivalent surface representation (what we could call a Digital Terrain Model). 
3.2. Chosen process 
A direct use of the DEM could have lead to an interactive process, by direct 
implementation of a control of the displacements in the stereo pair by an off line 
computed DEM. The study that was undertaken in 1991 aimed at going further, 
proposing a pre-computation of cartographic contour lines. The correlation of the DEM 
by the same kind of algorithm that is used at decametric resolutions [RIAZANOFF 92] 
was to be followed first by an automatic detection and suppression of isolated maxima 
(that could correspond to trees or buildings), and then by a smoothing procedure 
preserving the altimetric position of ridges and valleys. The operator would then edit 
the result to correct the remaining errors (dense vegetation areas, extended correlation 
failures, etc...). 
Up to now, the initial input is limited to the control parameters of the algorithms. If the 
final smoothing could benefit much of manual break line input, we didn't wish to make 
it interactive: 
- first, if it seems difficult to predict exactly which set of line will yield the best result, 
even the very best result will need some editing; making the operator interfere at too 
many stages of the process wasn't thought valuable; 
- secondly, global computation of the smoothing (taking into account all the constraints) 
shows more reliable than local corrections according to each given break line: the 
computational time for this operation prevents any interactivity. 
3.3. Evaluation 
Evaluating an automatic procedure is much more straightforward than evaluating an 
interactive one. If the data accuracy issue remains the same, in first analysis, the 
productivity issue is much more operator independent: 
- providing that the operator doesn’t interfere in the middle of the process, the 
computational time can be discarded (night batch processing); 
- neglecting all the operational data management problems, the capture time is restricted 
to the editing, and can be considered as a function of the number of errors (and of their 
extension).