ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision‘, Graz, 2002 
  
contrast to the shape-based matching. By all means, it should be 
pointed out that the computation time of PatMax® in this exam- 
ple is much slower than in the example above. Thus, in this case 
our new approaches are not only dramatically faster than the tra- 
ditional methods but also considerably faster than PatMax®. The 
reason for the totally different computation times of PatMax® in 
the two example sequences is the automatic computation of the 
coarse grain limit (see Section 3.2). During the first sequence us- 
ing the shifted IC the grain limit was automatically set to 3.72 and 
during the second sequence using the rotated IC the grain limit 
was automatically set to only 2.92. There is no obvious reason 
for this difference, because the object was the same in both cases. 
Experiments have shown that the automatic computation of the 
grain limit may result in a totally different value if the region of 
interest is shifted by just 1 pixel without changing the number of 
edge points within the region. 
4 CONCLUSIONS 
We presented an extensive performance evaluation of six object 
recognition methods. For this purpose, the normalized cross cor- 
relation and the Hausdorff distance as two standard similarity 
measures in industrial applications were compared to PatMax® 
— an object recognition tool developed by (Cognex, 2000) — 
and two novel recognition methods that we have developed with 
the aim to fulfill increasing industrial demands. Additionally, a 
new method for refining the object’s pose based on a least-squares 
adjustment was included in our analysis. We showed that our 
new approaches have considerable advantages and are substan- 
tially superior to the existing traditional methods especially in 
the field of robustness and recognition time. Even in compari- 
son to PatMax® particularly the shape based matching in com- 
bination with the least-squares adjustment shows very good re- 
sults. In contrast, we exposed some inconsistencies when using 
PatMax®: restricting the parameter space to translations causes 
the recognition rate to drop down dramatically. The automatic 
setting of the grain limits is problematical since significantly dif- 
ferent results are obtained using the same object and, although a 
high value for the grain limits leads to a fast computation it also 
results in a high risk of returning false positives. In contrast, a 
low value means higher robustness but slow computation. 
In most cases the shape-based matching approach and the modi- 
fied Hough transform show equivalent behavior. The shape-based 
matching approach should be preferred when dealing with in- 
tense illumination changes and situations where it is important to 
know the exact orientation of the object. In contrast, the modified 
Hough transform is better suited when either the dimensionality 
or the extension of the parameter space increases and the compu- 
tation time is a critical factor. 
The breakdown points of the normalized cross correlation are its 
low robustness against occlusions/clutter and non-linear illumina- 
tion changes. The relatively slow computation is another factor 
that limits its applicability. As a breakdown point of the Haus- 
dorff distance its trend to return false positives should be men- 
tioned, which often occur in the presence of clutter. 
Aside from these conclusions, it should be pointed out that some 
of the results might change if we chose, for example, other im- 
plementations of the approaches, other parameter constellations 
or other image sequences. Therefore, our comparison is more of 
a qualitative nature rather than of a quantitative one. Neverthe- 
less, our results are very objective and help potential users to find 
the optimum approach for their specific application. To facilitate 
an extended comparison including other recognition methods, in- 
terested parties are requested to send an e-mail to the authors in 
order to get the sequences that we used for the evaluation. 
A - 374 
REFERENCES 
Ballard, D. H., 1981. Generalizing the Hough transform to detect 
arbitrary shapes. Pattern Recognition 13(2), pp. 111—122. 
Borgefors, G., 1988. Hierarchical chamfer matching: A para- 
metric edge matching algorithm. IEEE Transactions on Pattern 
Analysis and Machine Intelligence 10(6), pp. 849—865. 
Brown, L. G., 1992. A survey of image registration techniques. 
ACM Computing Surveys 24(4), pp. 325-376. 
Cognex, 2000. Cognex MVS-8000 Series - CVL Vision Tool 
Guide. Cognex Corporation. CVL 5.5.1. 
Lai, S.-H. and Fang, M., 1999. Robust and efficient image align- 
ment with spatially varying illumination models. In: Computer 
Vision and Pattern Recognition, Vol. II, pp. 167-172. 
Matrox, 2001. Matrox Imaging Library - User Guide. Matrox 
Electonic Systems Ltd. Version 6.1. 
Olson, C. F. and Huttenlocher, D. P., 1997. Automatic target 
recognition by matching oriented edge pixels. IEEE Transactions 
on Image Processing 6(1), pp. 103-113. 
Press, W. H., Teukolsky, S. A., Vetterling, W. T. and Flannery, 
B. P., 1992. Numerical Recipes in C: The Art of Scientific Com- 
puting. 2nd edn, Cambridge University Press, Cambridge. 
Rucklidge, W. J., 1997. Efficiently locating objects using the 
Hausdorff distance. International Journal of Computer Vision 
24(3), pp. 251—270. 
Steger, C., 2001. Similarity measures for occlusion, clutter, 
and illumination invariant object recognition. In: B. Radig and 
S. Florezyk (eds), Mustererkennung 2001, Springer, München, 
pp. 148-154. 
Tanimoto, S. L., 1981. Template matching in pyramids. Com- 
puter Graphics and Image Processing 16, pp. 356—369. 
Ulrich, M., 2001. Real-time object recognition in digital im- 
ages for industrial applications. Technical Report PF-2001-01, 
Lehrstuhl für Photogrammetrie und Fernerkundung, Technische 
Universität München. 
Ulrich, M. and Steger, C., 2001. Empirical performance evalua- 
tion of object recognition methods. In: H. I. Christensen and P. J. 
Phillips (eds), Empirical Evaluation Methods in Computer Vi- 
sion, IEEE Computer Society Press, Los Alamitos, CA, pp. 62— 
76. 
Ulrich, M. and Steger, C., 2002. Performance evaluation of 
2d object recognition techiques. Technical Report PF-2002-01, 
Lehrstuhl für Photogrammetrie und Fernerkundung, Technische 
Universität München. 
Ulrich, M., Steger, C., Baumgartner, A. and Ebner, H., 2001a. 
Real-time object recognition in digital images for industrial appli- 
cations. In: 5th Conference on Optical 3-D Measurement Tech- 
niques, Vienna, pp. 308—318. 
Ulrich, M., Steger, C., Baumgartner, A. and Ebner, H., 2001b. 
Real-time object recognition using a modified generalized Hough 
transform. In: 21. Wissenschaft-Technische Jahrestagung der 
DGPF, Vol. 10, Konstanz, pp. 571—578. 
Wallack, A. and Manocha, D., 1998. Robust algorithms for ob- 
ject localization. International Journal of Computer Vision 27(3), 
pp. 243-262.