Interest Points Image Detectors: Performance Evaluation Perez Daniel Karina Ruby 1 , Escamilla Hernandez Enrique 1 , Nakano Miyatake 1 Mariko Perez Meana Hector Manuel 1 , Gabriel Sanchez Perez 1 1 National Polytechnic Institute, ESIME Culhuacan Santa Ana Avenue # 1000 Col. San Francisco Culhuacan Coyoacan, Mexico, Zip Code: 04430 krperezd@hotmail.com, eescamillah@ipn.mx, mariko@calmecac.esimecu.ipn.mx Abstract Most of computer vision applications employs interest points of the scene (images or frames) to undesrstand it. Therefore an accurate detection of the interest points is es- sential for many computer vision applications. The interest points must be invariant to rotation, zoom, blur, illumina- tion and point of view changes. This paper presents a per- formance comparison of the most popular interest points detectors, such as Harris, Harris-Laplace, Laplacian of Gaussian and SIFT, in order to know whch of them could be the most accurate. In the evaluation, the average vari- ation of the interest points through those changes is com- pared. Frame sequences of several environments are used to perform the evaluations. The comparison results can be used for selection of an adequate detector and furthermore improvement of the performace of them. 1. Introduction Nowadays computer vision becomes a very important scientific field which has a myriad of applications [1], such as 3D reconstruction, content-based image retrieval, object recognition, mobile robot localization, etc. Most of those applications are based on the detection of interest points in images or video frames. Therefore many researchers have been concentrated on developing efficient detections algo- rithms or improving the robustness of conventional algo- rithms; as examples, the works by Moravec [2], Harris and Stephen [3], Mikolajczy [4], Lindeberg [5], Lowe [6] can be mentioned. A brief description of some of these detectors is done in the next section. To employ efficiently the interesrt points in above men- tioned applications, the detection should be reliable and then the number of detected point and point’s position must not vary under several situations, such as changes of point of view, zoom, illumination. However interest points detected by many detection al- gorithms [3, 4, 5, 7, 8] still present some instability, depend- ing on the variation, the number and position of the interest points may vary. So it is important to analyze the perfor- mance of several detectors in order to know which detector works better under some specific conditions [8, 9]. In this paper, the performance evaluation of some inter- est point detectors is carried out using the repeatability of interest points as evaluation criteria. The results of this eval- uation lead the way for future improvements. The rest of this paper is organized as follows, in Section 2 a descrip- tion of four interest point detectors used for evaluation is given, in Section 3, the performance evaluation results are provided and finally in section 4 the conclusions are done. 2. Feature Extraction In this section a brief description of Harris detector, Harris-Laplace detector, Laplacian of Gaussian detector and SIFT detector, as the most popular interest point detectors, is given. 2.1. Harris Detector Harris Corner detector (HD) is a widely used detector, which is based on Moravec [2] detector. In HD, all possible small shifts can be covered by (1) unlike to Moravec detec- tor where only a discrete set of shifts at every 45 degrees are considered E x,y =  u,v w u,v (I x+u,y+w - I u,v ) 2 (1) where I u,v denotes the image intensities at the coordi- nate (u, v) and the function w u,v is a rectangular Gaussian response window centered at (u, v). For small shifts [u, v], the equation (1) can be written in a equivalent form as (2). 2011 Electronics, Robotics and Automotive Mechanics Conference 978-0-7695-4563-9/11 $26.00 © 2011 IEEE DOI 10.1109/CERMA.2011.29 137