5 NJ T € €»
oO ^ 0G...
1C
al
Shih-Hong Chio
straight lines, for horizontal roof patch only horizontal lines. The semantic information of a line segment is composed of
the height information of the two endpoints of a 3-D line segment and the average height of that segment.
Then, the operator verifies whether this semantic information is correct or not and whether this 3-D line represents the
border of a roof patch or not. In case that this 3-D line is denied by the operator, the system will prompt the next one.
Only if the initial guess of the roof patch model (the semantic information) is wrong, the operator can change it.
Otherwise, the operator will confirm it and the system guides the operator to the next step. In the next step, the operator
has to select relevant line segments associated to the initial 3-D line to form the roof patch. The selection is done mono-
scopically by clicking the cursor onto the lines. Without losing the generality, we will assume that in the subsequent
interactive procedure all the selection and confirmation are done in the left image of the stereo pair. Note that although
the selection is done in one image, but the selected associated line segments could be labeled as 2-D or 3-D lines
depending on the results from the previous stage of linking and matching in object space
Since no feature-extraction method can guarantee a
Oblique Line complete extraction of linear features, the associated line
— segments are in general only piecewise and represent only
fragments of roof edges. Sometimes there is even no line
for selection at all. If the system finds out that the
| information aggregated up to now is still not enough for
Horizontal Line the reconstruction of that roof patch after the selection of
3 : the associated line segments, it will request the operator to
Horizontal Roof Patch Slope Roof Patch | position those corner(s) which is (are) necessary for the
reconstruction of that roof patch. In subsequence we will
Fig.2. Plane Model for Roof Patch Reconstruction and call these corners the "lacking corners". The locating of
Semantic Information of 3-D Line these lacking corners is done mono-scopically in the left
image by clicking the cursor on the corner position. The
system then search in the Point Database of the left image to find a point which is probably a corner point at that
location. If in the vicinity of the cursor position there is no point in the Database, the system simply takes the current
cursor position as the corner point.
Depending on how many lines and lacking corners are selected and how much information is inhered in the lines and
corners, different situations could arise. Fig.3 illustrates 8 possible cases after the selection of associated lines and
positioning of the lacking corners. Assuming that for all cases the lower side edge (displayed in thick lines in Fig.3) is
the initially confirmed 3-D edge segment of a roof patch. The associated lines selected by the operator are displayed as
thin lines. The dotted lines mean missing lines (either there is no line at all for selection or can not be confirmed by the
operator). An unmarked corner means it is determinable at least in 2-D image space by the system itself. A double circle
represents the lacking corner.
The aim here is to obtain the 3-D coordinates of all four corners of the roof patch. But since there are so many different
situations, there is no unique way to obtain them. In the following we will see that for each different case the solution is
also different.
The most ideal and simplest case would be
/ ; QUE jum 5 | the upper left one shown in Fig.3. In that
5 : case, for all four edges, some segments of
f the roof edges have been confirmed by the
operator. If all segments are 3-D lines, 3-D
coordinates of all corners can be calculated
directly through intersection of two
Benn © e--———— 6 9——— © © © adjacent line segments pointing to the
/ f / Fari i j corner. For all other cases the 3-D
coordinates of the corners have to be
; determined indirectly through information
© provided by the geometrical structure
shown in Fig.3 and the semantic
Fig.3. The Relation of Selected Lines, lacking Corner(s) within information associated with the selected
One Complete Roof Patch lines and the roof patch model.
Thick Line in Every Quadrangle: First Prompted 3-D Line In the analytical Photogrammetry we know
Thin Lines: Selected Lines that a point in space can be determined in
Double Circles: The Lacking Corners two ways. The first way is using stereo
Photogrammetry by forward intersection of
two rays from two stereo images. The second way is using the analytic of single image by knowing at least one of the
space coordinates, usually the height Z, of that point. Both methods will be used here depending on the available
information. If the height information could be derived through any way, we will use the second method. Otherwise we
6. m Í © © ®
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 185
