2.3 Pose estimation and viewpoint projection 
We use EPnP (Lepetit et al., 2009) to estimate the pose of 
image I in the reconstructed space. The calculated 3D 
coordinates X(I) is then projected into images covering the 
nearby environment. The viewpoint of image I is determined 
when p'(I) lies in effective area of an image plane. 
3. RESULTS AND DISCUSSION 
To test the presented method, we firstly took the photo shown in 
Figure la. And the actual place where we took the image is 
marked with a red circle. Then we went to the square and took a 
collection of overlapping images (Figure 2). To guarantee the 
accurate 3D using Bundler, we keep relatively small angle 
between viewpoints of neighboring images. The reconstructed 
3D scene is illustrated in Figure 3, in which the position of 
3704 points and pose of 24 cameras are visualized. Figure 4 
shows 12 of all 3704 points and their projections in an image. 
These projected points are transferred to image I in Figure 5. 
Figure 6 illustrates the actual position and projection of the 
calculated 3D viewpoint of image I to an image covering the 
building. Figure 7a, 7b and 7c give show the presented method 
tested another dataset. 
The computation of our method is mainly cost by image 
matching procedure. Not all of the cameras can be registered 
using Bundler, and sometimes the reconstruction is not accurate. 
The result of point transfer has many outliers which often lead 
to fault estimation. In this research we cut off some outliers by 
hand and recalculate the viewpoint of image I using EPnP. 
Figure 3: The reconstructed 3D scene (point cloud) 
s BE RR 
Figure 4: Reconstructed points projected to an 1mage (red 
crosses) 
  
Figure 5: Point transfer (green crosses) 
Figure 6: The estimated viewpoint (green point) of image I and 
its actual place (red point) 
  
Figure 7a: Anoter test image