International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
Test data 2 
Area with high 
buildings. 
Size: 3000*3000 
  
Test data 3 
The normal area 
ly Size: 4096*4096 
    
Figure 5 three test data set. 
The time cost of each processing step is showed in table 1. 
Table 1 processing time 
  
  
  
  
  
Processing step Spending time (ms) 
Datal | Data2 | Data3 
1. Wallis filter 672 1321 |. 2321 
2. feature extraction (Harris) 122 231 321 
3. calculate left image vector | 32 35 37 
(10*10 points) 
  
4. calculate right image vector | 753 1424 | 2531 
(all pixels) 
  
  
  
  
  
  
  
  
5. find potential correspondences | 12 12 11 
6. remove error points by GHT 11 11 12 
Total 1602 | 3034 | 5233 
  
Table 2 processing time by SIFT 
  
Spend time (ms) 
  
Datal Data2 Data3 
  
  
  
  
  
Total 42304 93872 244387 
  
  
  
(Download code form 
http://www.cs.huji.ac.il/~ofirpele/SiftDist/code , 
Compiling by Microsoft Visual V++ 6.0 ) 
The matching results are showed in Figure 6. 
     
   
  
  
Figure 6 Matching Result of The Method 
As can be seen from the results, the image matching 
algorithm based on the rotation vector field can indeed solve the 
problem of image rotation , and the computing speed is much 
faster than the traditional template matching algorithm in which 
cach pixel is involved in complex operations ( such as 
convolution). 
As we can see, the method can achieve good matching points 
which can be used lately as the initial-value for the accurate 
matching. 
4. CONCLUSION 
According to the above experiments, conclusions can be 
drawn as following. 
Firstly, the matching algorithm based on rotation vector field 
is an optimized method which greatly reduces the computation 
and improves the matching speed. 
Secondly, the algorithm can get good results in regions with 
rich features and few repeated features. And it’s not suitable for 
regions that with poor feature or have a large number of 
repeated features. 
Thirdly, the accuracy of the matching points obtained by this 
method is not that high. And these points can’t directly 
participate in photogrammetry processing but can be used as the 
initial-value for the accurate matching. 
Our further research will focus on the scale-change problem 
in the image matching which hasn’t been solved yet. 
REFERENCES 
DAVID, G. LOWE. 2004. Distinctive Image Features from 
Scale-Invariant Keypoints. International Journal of 
Computer Vision, Vol. 60, No. 2, pp. 91-110 
Herbert Bay, Tinne Tuvtellars and Luc Van Gool. 2006. SURF: 
Speeded Up Robust Features. European Conference on 
Computer Vision. pp. 404-417 
Zhang Li, Zhang Zuxun,Zhang Jianging. 1999.The Image 
Matching Based on Wallis Filtering. Journal of Wuhan 
Technical University of Surveying and Mapping, Vol, 24 
No.1 ,pp. 24-26 
Sim D G, Kim H K, Oh D I. 2000. Translation, scale, and 
rotation invariant texture descript or for texture based image 
retrieval .IEEE Int Conf Image Process, No.3, pp. 742- 745 
Farhan Ullah, Shunichi Kaneko. 2004. Using orientation codes 
for rotation invariant template matching. Pattern 
Recognition , 37, pp. 201- 209 
Yan Ke, Rahul Sukthankar. 2004. PCA2SIFT: A More 
Distinctive Representation for Local Image Descriptors. 
Proceedings of the 2004 IEEE Computer Society 
Conference on Computer Vision and Pattern Recognition, 
pp. 506-513 
120