ORGANISMOS DE SOCORRO LOCAL

Point clouds M najdra (Figure 5.26) and T arxien (Figure 5.29) acquired from two Maltese CH sites are used in this section of the evaluation, in addition to a point clouds acquired from a stone church in the town of Lans le Villard, France.

M najdra represents a section of the Mnajdra pre-historic temples site that was acquired in 2005 by Heritage Malta. The modelled surface precision was stated as +/- 2mm. Due to considerable stone erosion there are hardly any smooth surfaces present. In this case study, segmentation is used to discriminate between the various stones composing the temple apses. Figure 5.26 shows the point cloud rendered after point type labelling (top) and following RanSaC plane fitting (bottom). The image shows all the planar surfaces (rendered using different colours) fitted over all the different surface segments of the Mnajdra temple. The partitioning induced by RanSaC is efficient and reliable due to the fact that planes are fitted on subsets of related points (produced by the region-growing process) rather than the whole data set. The 98% of the surface points are fitted to 930 plane primitives. Without prior segmentation, RanSac does not converge properly and only two planes (see Figure 5.28 top right corner) are fitted to the floor of the temple in approximately 4 minutes. Moreover, each of the two plane primitives cover points which are found on different, unrelated parts of the temple. PaRSe outputs a set partition with geometric planes fitting the data in a very accurate way with the general structure of the temple clearly visible.

Two queries are carried out on the resulting structure graph to extract the walls and the floor of the temple. Figure 5.27 illustrates the results returned by these two queries. In the first instance, three seed segments are used in order to return the three components of the temple. Note how the query search follows the rubble wall until megalith stones positioned perpendicular to the wall are encountered. Figure 5.28 shows the largest surface and edge segments resulting from the region growing process. As would be expected the largest surface segment consists of points sampled from the floor of the main part of the temple, whereas the largest edge segment is made up of points connecting the entire rubble wall structure. The middle row provides a closer view (of part) of the rubble wall present in the site. Around 35 stones are automatically identified in this part alone. The edge segment contributes towards the partitioning of the rubble wall into a number of surface segments representing the individual stones making the apse.

Seed Segment

Seed Segment

Figure 5.27: A query graph (using seeds) is used on segments from each apse wall to return the internal walls of the temple is shown on the left. Another query is used to return the points making up the main floor (within the apses) of the temple.

Figure 5.29 shows the segmentation results obtained for the Hal-Tarxien tem- ples. As in the case of Mnajdra, the Hal-Tarxien data set was made available by Heritage Malta. 728 surface segments are directly planar segments and an addi- tional 109 segments are split into *·planar segments, resulting in a total of 977

*·planar segments. Figure 5.30 shows the main structures within the Hal-Tarxien pre-historic site, with the apses and floors of both temples easily identified on the left-hand side as surface segments. The right-hand side image shows the edge segments, which effectively provide an outline of the temple structure. Fig- ure 5.31 zooms into two particular features of the site, a cylindrical bowl and stairs. In the first case, the bowl’s points falling on the irregular top structure are represented by a unique surface segment resulting from the region-growing process. In the second case, 20surface·planar segments represent a series of steps inside the site. Each step is represented by approximately 100 points, totalling around 0.07% from a total of 2.7 million points covering the site.

The region-growing phase of PaRSe is essential to the successful extraction of meaningful elements in the site. Figure 5.32 demonstrates the results obtained (using the same parameters) when RanSaC plane fitting is directly applied to the point cloud. The largest segments consists of 281K points and consisting of several points sampled from unrelated parts of the site. The resulting set partition (only the largest 6 segments of 148 are shown) is not useful in distinguishing between the different components of the site and would certainly require a CH professional a considerable amount of time to work with.

Figure 5.33 illustrates the point cloud of a patrimonial stone church in the town of Lans le Villard, France. The raw point cloud is available on the Aim@shape

Figure 5.28: Top left images illustrates the largest surface (green) and edge (black) segments resulting from the region growing process. Top right image illustrates the results obtained when RanSac plane fitting is carried outwithoutprior region-growing segmentation.

3 Hal-Tarxien Temples Point Cloud

Position Data

837 Surface Segments after Region Growing

977 *.planar segments

after RanSaC Fitting

Figure 5.29: Hal-Tarxien point cloud; top illustrates raw data, middle illustrates seg- ments produced after region-growing process, and bottom illustrates segments pro- duced after RanSaC fitting process.

Figure 5.30: The main structures within the Tarxien pre-historic site. Left-hand side image illustrates the surface segments, whereas the right-hand side image shows the edge segments. Edge segments effectively provide an outline of the temples.

Figure 5.31: Top left illustrates a view of the Tarxien point cloud. This same view is shown on the top right hand side, this time showing the various surface segments produced. The bottom row zooms on two structures in the site. The bowl like structure is shown in the left hand side, with different segments highlighting its shape, whereas the right hand side illustrates segments representing a flight of steps.

281K 164K 127K

102K 100K 93K

Figure 5.32: Top image illustrates the 148 segments extracted, whereas the second and third rows illustrate, individually, the largest six segments produced using the traditional RanSaC plane fitting process, segments which are clearly not very useful in term of describing the elements making up the site. Numbers within each box indicate the number of points in that specific segment.

website (Falcidieno, 2004) and was scanned with a Leica Cyrax scanner in 2006. The bottom row shows the raw point cloud (left),edge·* (middle) and surface·*

segments (right). The top row shows a photograph of the church and three

surface·planar segments extracted from one surface segment produced at the region growing phase. A considerable number of points are labelled as edge, es- pecially on the roof on the church. This is evident in Figure 5.34 with the edge segment outlining the stone slabs making up the roof. This edge segment is useful in extracting individual stone roof slabs as shown in the bottom row. The site contains a number of tombstones directly adjacent to one of the walls. Figure 5.35 shows the extracted segments representing these tombstones. PaRSe creates two segments representing the face and side of one, whereas the other, due to considerable noise is over-segmented, and consists of 9 surface and edge segments. As shown in the same figure, edge segments representing window rails are also automatically extracted. Additional site details are shown in Figure 5.36 includ- ing the cross at the top of the church as an edge segment. Over-segmentation is clearly visible on some sides of the church tower resulting from a large number of points which are labelled as edge.