En relación a la actitud de los jóvenes universitarios de la

15 (pixels)

Figure 10.8; The y-disparity distribution (left) resulting from the PM F w ith m atching strength hmit. T he edge threshold was 0.5/C. The m anually deter­ m ined true disparity image (right).

a tem perature of 0 .5 if. Each pixel was displayed as a 5 x 5 square because the m atched pixels are sparsely distributed. Many falsely m atched pixels w ith large disparities are shown on the lower part of the image where only low stratus clouds exist.

10.3.2

Area-based M atching

An area-based m atching software called G OTCH A was used to m atch the test im ages. The G OTCH A is a powerful area-based m atching algorithm which was developed in Photogram m etry and Surveying Departm ent in U niversity College London. It is based on the A daptive Least Squares Correlation (ALSC) [Gru85] w ith a region growing technique which allows autom atic m atching of the w hole image from a small number of seed points [OC89] (See also A ppendix F .2). This m atching algorithm showed very reliable and accurate m atching results for the im ages with continuous disparities such as topographic im ages [DM89].

For the GOTCH A m atching of the cloud images, the m atched pairs from the P M F results were used as seed points. The seed points do not necessarily

1 0 .3 Ed g e a n d Ar e a Ba s e d Ma t c h i n g o f Cl o u d Im a g e s 1 5 9

have to be very accurate because falsely matched seed points are automatically rejected by the ALSC algorithm. In addition, the disparities of the seed points are accurately refined by the ALSC before the region growing begins. Even seed points with random disparities have been apphed to the GOTCHA successfully [OD92].

Out of the most which determines the matching performance of the GOTCHA is the patch size. The bigger the patch size, the more data is used for the matching of one pixel, and hence more computations are required. A larger patch may imply more accurate matching because more information is used to measure similarity. However, a large patch contains large dispar­ ity variations, or possibly, large disparity discontinuities, which result in poor matching performance. Therefore, the patch size must be chosen for different types of images.

The GOTCHA finds m atch estimates at the points of intersection of a regularly spaced grid in the nadir image. Therefore, the num ber of matched points in the GOTCHA depends on the size of the grid on which the matching is performed. In general, a finer grid provides more m atched points but it also requires longer processing time.

The results of the GOTCHA matching are shown in Figure 10.9. Four different patch sizes were selected in order to study the dependency of the patch size on the matching accuracy. For each patch, different grid sizes, 2x2, 3 x 3 and 4x4, were applied. As shown in the graph, smaller patches resulted in more accurate matching. This is due to the disparity discontinuities at the cloud object boundaries. The GOTCHA failed to m atch pixels when the patch radius was smaller than 3 because not enough information was contained in the patch. The matching accuracy did not depend much on the size of the grid, and hence, a smaller grid pitch gave the larger num ber of matched points at the cost of a longer processing time (details on the processing time of each algorithm are discussed in Section 10.5). The overall y-disparity RMS error is approximately 3 pixels which is much more accurate th an th e PMF-with-MSL as shown in Figure 10.9.

Figure 10.10 shows the y-disparity distribution of the GOTCHA m atch­ ing with the patch radius of 5 pixels and the grid size of 2x2. Compared with the true disparity in Figure 10.4, the GOTCHA provided good matching

1 0 .3 Ed g e a n d Ar e a Ba s e d Ma t c h i n g o f Cl o u d Im a g e s 1 6 0

6.0

A ---A patch_radius = 3 □ ---□ patch radius = 5 O---O patch radius = 7 O---O patch radius = 9 5 .0 ■O 3 .0 2.0 3 0 0 0 0 4 0 0 0 0 10000 20000

Num ber o f m atched clou d pixels

Figure 10.9: The y-disparity RMS error of the GOTCHA algorithm showing accuracy as a function of grid pitch (2,3 and 4).

except th at it overestimated the area of the high cirrus cloud field and it re­ sulted in false m atching for isolated high cirrus cloud objects. This is because an area-based m atching algorithm does not m atch the regions with disparity discontinuities accurately.