The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
DEM, but for an irregular geometry body like a building and its 
corresponding appearance attributes, for instance, the texture 
image. This work is still not satisfactory, and it always needs 
plenty of human-computer interactions to get a content result. 
1.2.1 Simplification Considering Semantics 
Based on web 3D GIS environment, this method embeds vd as 
the characteristic dominant value in vertex attributes by Quadric 
Error Metric (QEM), and it also translates the semantic inhibit 
into the simplification for 3D object by means of figure 
distilling. Combined with the model’s semantic characteristic, it 
has well kept the model’s spatial characteristics in the process 
(Coors, 2001)This method has imported the concepts of focus 
structure and graphical abstraction to separately describe a sort 
of expressing way and the semantic simplification, the former 
has the most visual importance in the objects and the latter can 
converse the semantic restricts into 3D objects under a 
consideration of the background environment. 
But this algorithm couldn’t directly create each building 
model’s focus structure, because it may differ among every 
object. Furthermore, the graphical abstraction may be difficult 
to accord with the description of the focus structure, which 
means what details should be saved is still not solved. Another 
drawback is the neglect of the impact from the building’s inner 
complex structure in its simplification method and process. 
1.2.2 Simplification Considering Features 
Martin Kada put forward a simplification algorithm aiming at 
3D city building models, in which three kinds of characteristics 
like co-planarity, parallelism and uprightness among surfaces 
are detected and adopted by the edge collapse to simplify. This 
method can well keep the shape and relational surface attributes 
after removing some of the building’s characters such as texture 
(Kada, 2002, 2007). 
Its drawback is obvious: Firstly, the experimental model is so 
simply that it couldn’t roundly represent the building model’s 
complexity. Secondly, the drawback of co-planar contracting is 
still unsolved, thus it is not fit for the simplification of the 
complex building model in VGEs. Thirdly, the process is 
strongly restricted by the minimum measurement of the original 
building model’s components, and then its ability of reduction 
is so limited. Furthermore, its simplifying ability is really 
limited neither for arbitrary fold meshes nor for unfold meshes. 
Besides, Thiemann has divided the model according to 
characteristics checked out and express them in the CGS way. 
And then, he made the simplification operation through the 
CGS tree (Thiemann, 2002; Thiemann et al., 2004) Its 
excellence is that simplification is on consecutive scale and also 
owns the possibility for semantic expanding. But considered as 
a method for generalization, it couldn’t combine the 
neighbourhood buildings (Sester, 2007). 
1.2.3 Simplification Based on Scale Space 
Aiming at the building model data, Andrea Forberg brought 
forward a 3D model simplification method based scale space 
(Forberg, 2004, 2007; Forberg et al., 2002). Its main academic 
foundation of is the mature scale space theory, including 
mathematic morphology and curvature space theory. It has used 
different form operation such as Erosion, Dilation, Opening and 
Closing to control the unit and separate among different parts. 
This method is developed and realized on ACIS 3D Geometric 
Modeller and VRML. 
The simplification method based on scale spatial theory gets its 
prodigious limits in simplification. For instance, it sometimes 
needs other means to deal with no-orthogonal characteristics 
like housetop, including rotating, correcting and so on (Sester, 
2007). Besides, this algorithm can’t assure the maintenance of 
the building’s characteristics, and it neglects consider some 
attribute contents like material and texture. 
Facing with the models, in summary, the traditional 
simplification methods have three drawbacks. Firstly, methods 
above can not accurately locate the portions needed to be 
simplified. Secondly, the most traditional simplification 
methods in computer graphics are restricted to simplify a single 
3D object with continuous surface mesh but not a set of meshes, 
but a complex building model is just a set of mesh because each 
component is a mesh. Thirdly, due to the simplification 
methods do not take human perception information to drive the 
simplification operations, so that the simplified results can not 
accord with the rule of human perception, and the LOD models 
derived from the simplification methods are difficult to ensure 
the continuity on visual effect. 
2. PERCEPTION-DRIVEN SIMPLIFICATION 
FRAMEWORK 
2.1 Perceptual Details of 3D Building Models 
Two choices are given to us for exploration of perceptual 
details of 3D building models: one is to do the detail analysis 
based on a geometry definition of an object, such as analyzing 
the geometrical characteristics represented by the building 
model’s geometrical primitives (vertex, edge, triangle, or body, 
etc.); the other is to enter on the object’s 2D rendering image. 
Both these methods own their excellences and shortcomings. 
Geometrical Characteristics: Now, these 3D models are 
expressed in the form of geometrical primitives, after which it is 
possible to compute various geometrical details of the 3D 
model, such as curvature, length. 
But, any geometrical details can not avoid the problem. A 3D 
model is not only totally expressed by geometrical primitives, 
but also by textures reflected to the material attributes on the 
model’s surface. For those geometrical models with textures, 
the textures contain a great lot of visual information having a 
deep influence on the perception, while the pure geometrical 
details’ calculation entirely neglects these visual details. 
Rendered Image: It is possible to gain a model’s rendered 
image by setting a camera and its parameters and pre-choose a 
rendering mode. Because the human beings are acquiring an all- 
sided perception directly by looking into the rendering images 
using their naked eyes, it is possible to extract the 3D model’s 
perceptual details by detecting the situation from human eyes to 
models reflected by the model’s rendered images. The rendered 
images have more accurately and integrally reflected the 
information on the model’s surface, such as its figures, natures, 
textures, illuminations, all of which are factors having a deep 
impact on human perception. And the integrity of information 
laid a foundation for the model perception.