651 
ARCHITECTURAL SCENE RAPID RECONSTRUCTION BASED ON FEATURES 
Y. Ding a,b- ‘, J. Q. Zhang b 
institute of Earth Observation and Space System Delft University of Technology 
Kluyverweg 1, 2629 HS, Delft, The Netherlands -Yi.ding@tudelft.nl 
b School of Remote Sensing Information Engineering Wuhan University 
129 Luoyu Road, Wuhan 430079, China 
KEY WORDS: 3D reconstruction, feature matching, differential evolution, visualization, robust estimation 
ABSTRACT: 
Retrieve the structure of model and the motion of camera is a classical and hot topic in computer vision and photogrammetry. A lot 
of automatic or semiautomatic techniques have been developed to optimize the retrieving processing from accuracy, stability and 
reality perspectives. These techniques are variant from data source, feature selection for matching, feature clustering and 3D model 
representation. The optimization algorithm and a completely automatic system are still under exploring. In this paper, we use some 
image-based algorithms for feature selection and matching of 3D man-made scene reconstruction. We present a robust point 
matching algorithm with RANSAC estimator, and compare two methods of line matching in a complex man-made environment. We 
point out the degeneracy when use epipolar line as a constraint to match line, instead use a global optimization method. Our 
experiments show that the proposed method is robust in a complex man-made scene. 
1. INTRODUCTION 
Observing the world through the eyes of machines or imitating 
the human ability of perception is the essence of computer 
vision and photogrammetry. Computer vision is an integrated 
subject of computer technology, image processing, pattern 
recognition, and computer graphics. There are many interesting 
research areas in computer vision, for example, robot 
navigation, traffic tracking, face recognition, recovering 3D 
structure of the environment, and so on. Among these research 
areas, recovering architectural 3D models from complex man 
made environment has been a very hot one in recent years. 
Applications of architectural 3D models are, for example, 
tourist navigation and heritage protection. There exist a wide 
variety of methods related to model recovery (Baillard, 2000 
and Cantzler, 2002). The differences between these methods are 
from the manner of getting data to the representation of the 
scene model. The most appropriate model recovery method 
depends on the type of a scene that is to be reconstructed, and 
the application requirements like real-time/unreal-time, detail 
representation degree, and etc.. 
Figured is a general flow chart describing the processes of 
model reconstruction. Most of the existing image-based 
methods are based on three-stage process (Bartoli, A., 2007). 
First, sparse features (points, lines, ect.) are extracted and 
matched in multi-view and then 3D features and camera pose 
are reconstructed by using structure from motion techniques. 
The remaining two stages are scene model selection and 
parameter estimating. After these steps we can then represent 
the model in the form of a dense depth map, triangular mesh or 
a set of space planes. 
When considering man-made environment, we choose distinct 
points to recover the camera motion and then to achieve plane- 
based model by clustering reconstructed line features. This 
plane-based model is motivated by reasons like very * 
constrained, compact representation and modify the 
reconstruction easily as described in (Bartoli, 2007; Hartley, 
2000). 
Figured Flow chart of reconstruction. 
Real man-made scenes are usually very complex with occlusion 
and noise. In this paper, we combine several algorithms to 
present a stable and automatic processing for recovery the 
structure of a scene and the motion of the camera with distinct 
features of point and line. We present a robust point matching 
algorithm with RANSAC estimator, and compare two methods 
of line matching. We point out the degeneracy when use 
epipolar line as a constraint to match line, instead use a global 
optimization method. 
The rest of the paper is organized as follows. Point feature 
extracting and matching to recover the motion of the camera are 
described in section 2. The line extracting and matching 
approach with constraints are discussed in section 3. 
Conclusions and future work are presented in section 4. Each 
* Corresponding author.