Geometry:
Resampling has a twofold effect: It degrades spatial resolution and
reduces locational (measurement) accuracy (Gp = A:3.46). Gridding
of the external information into image-raster does not necessarily
cause significant loss (but it is quantifiable). Geometric correc-
tions for deterministic errors can be accurate enough. Locational
errors after pre-processing can be analyzed by means of statistical
tools.
111,4 Image match
At matching, pictorial and geometric data are merged, which makes
theoretical assessment of matching accuracy very complex. The cor-
responding quality indicators depend on the matching technique [4].
In the case of a sequential search the common indicators are the
maximum of similarity assessment function (i.e., peak of correlation
function) and its maximum slope (at the point of inflection).
Matching accuracy depends also on the step size between successive
trials, and, further, whether or not a curve is fitted to discrete
points (i.e., results of trials) and the maximum is determined by
calculation.
If the least squares fit is applied, then the quality of match is
indicated by the descrepancies in intensity [1] or in location [3],
and the corresponding standard errors.
Overall accuracy of matching can be assessed from the statistics of
single matches involved, e.g. hystrograms and standard errors.
If resampling is applied iteratively during matching, image data are
degraded gradually, which can affect adversely the matching accura-
cy.
Reliability depends strongly on the input images, matching strategy,
acceptance threshold and the a-priori exclusion of anomalous
regions. Reliability is quantifiable in statistical terms (e.g., as
percentage of successful matches).
Time efficiency is not essential in off-line systems, though it is
not negligible. It can be expressed by cycle-times per trial, per
match, per full patch, and/or per full model.
111.5 Parallax data
Parallaxes are the gemometric by-product of image matching [4]. The
corresponding quality indicators are parallax accuracy of individual
points, parallax fidelity of a neighbourhood (i.e. a parallax "pro-
file" or a "surface"), reliability, and time-efficiency.
Parallax accuracy is the most important criterion for performance
assessment. It depends on the accuracy of matching segments and on
definition of representative conugate points in the segments (annex
3), geometric distortions (caused by model deformation), and on
thresholding at data compression rot
Indicators for parallax accuracy are standard parallax error
(annex 6) and error distributions (spatial, statistical).
Accuracy can be upgraded by collective processing of local parallax
data. The fidelity of such data (e.g., a parallax profile or sur-
face) depends on the interval between the adjacent (representative)
points, overlap of the adjacent image (target) segments, and the
smoothing filter (if applied).
348