The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
696 
which lie in frustum view region will be loaded, and search 
those tiles is by grid index (see figure 3a). If viewpoint moves 
over an adjacent tile the algorithm will tend to maintain a square 
of tiles centered on this new tile (which will be load in client 
memory new and becomes the current tiles). At the same time, 
the algorithm will remove some far tiles which are not within 
the field-of-view in order to free memory for the fetching of 
new tiles (see figure 3b). As figure 3a indicated, most of the 
memory of mobile device was consumed at the step. Note that 
the algorithm implicitly handles the case where viewpoint 
jumps to a new tile that is not adjacent to the current one. The 
quad-tree representation of tile data enables very fast view 
frustum culling. 
Figure3. Tiles management and adaptive loading 
a) . A square area centered on the viewpoint. 
b) . Square area preservation on viewpoint move 
a. 
b. 
Figure 4: Eliminating cracks, a) Cracks appear on tiles borders 
when adjacent tiles have different levels, b) Using dividing 
compulsively technique, crack effects are eliminated or 
attenuated. 
Farther in more, considering the tiles are made of a set of 
triangles, the triangles of the tiles will be judged one by one, 
and those trigngles which lie in the field-of-view are fetched. At 
last those triangles that lie out of view frustum will be removed. 
4.3 Tile rendering and cracks eliminated 
When the resolution levels are different between two adjacent 
tiles, there will be gaps on tiles boundaries which create a very 
unpleasant visual effect (showing in figure 4a) Classical 
approaches [Larsen and Christensen 2003; Lossa and hoppe 
2004] consist in modifying the geometry of the tile’s border by 
introducing new vertices and edges. Another classical method 
called filleting, introduced by Sun and also implemented in the 
NASA’s World Wind remarkable earth viewer is to add a band 
of vertical triangles around the edges of each tile. This band is 
stretched down to the lowest terrain elevation. Each side of the 
band is textured by stretching the corresponding line/column of 
texels. This scheme is fast but quite disgraceful to see for the 
user, especially if a neighbour tile has not been load yet. 
5. RESULTS AND CONCLUSION 
This paper proposes an adaptive approach to render large 
terrains. By this means, it simplify the scenes’ complexity 
efficiently and reduce the amount of data and therefore of 
graphical primitives to render in real-time. 
5.1 Experiment and results 
Our experiment selected Hp2110 as the experiment platform, 
and its primary configurations are shown in table 1. 
CPU 
Intel PXA270 
Memory 
64 Mb (32Mb RAM + 32Mb ROM) 
Screen resolution 
320X240 
Storage 
1GB SD card 
Operating System 
Pocket PC 2003 
Table 1. Configurations of Hp2110 
In this paper, another method which divides the triangles which 
are in coarse tile compulsively (see figure 4b) is proposed. If 
there appear cracks for the different resolution levels between 
adjacent tiles, the triangles which lie on the boundary in coarse 
resolution tiles will be divided compulsively to suit the fine 
resolution tiles. Even if this solution is not perfect, it is fast, 
simple to implement and gives satisfying results most of time. 
The experiment first measured the difference between the 
simplify scene and primal scene. The amount of primary 
triangles in client memory is about 70,000 (see figure 5a), and 
after being simplified, the amount of triangles reduced to 16064, 
almost 77.1 percent triangles are removed or cached (seeing 
figure 5b).