Scale Invariant Feature Transform Evaluation in Small Dataset Aeyman Hassan School of Computer Engineering University of Zawia Libya Email: a.hassan@zu.edu.ly Abstract—This paper investigates how we can achieve object recognition in an image by looking at some examples of training images. Scale Invariant Feature Transform (SIFT) is one proposal method to detect features in an image and then can use those features to distinguish between different objects. Therefore, my aim was to implement SIFT code to do recognition tasks using simple thresholding and evaluating this algorithm to find its strength and weakness points for a small dataset. The challenge here is to find the best threshold for examples of training images, which can work properly with query images. Keywords—object recognition,thresholding,evaluation,SIFT,small dataset I. I NTRODUCTION The features extracting and matching consider being a prior step in many computer vision applications. Finding corresponding matching features between two images is a crucial task to achieve image mosaic, camera calibration, ob- ject recognition. . . etc. Therefore, we have to find appropriate features which can fulfill our requirements. One of the most widely used methods for feature detection is Scale Invariant Feature Transform (SIFT). It is an algorithm to find and describe interest points in images (local features). SIFT was published by David Lowe in 1999. It is called invariant because it is invariance to illumination, scale, and rotation [1]. This algorithm can efficiently find feature matching even in exciting of a large database of local features. Tao et al. [2] presented performance evaluation for both SIFT and extend SIFT descriptors for object recognition. They applied several factors (Recall-precision, average precision, re- peatability rate and Receiver Operating Characteristics (ROC). Khan et al. [3] studied the efficiency of SIFT and Speed Up Robust Feature (SURF) with a various dataset. The criteria which were used in their work were matching accuracy and computation time. A comparative study between SIFT and SURF was introduced by Panchal et al. [4]. Grabner et al. [5] proposed an approximation of SIFT. They worked on increasing the speed of SIFT by using integral images to decrease the cost of scale space finding. In this paper, a study of SIFT evaluation was carried out in a small dataset by using simple threshold. We can see an overview of SIFT algorithm in section II. Presenting of the database which was used in the evaluation is in section III. Section IV discusses the experiments. Discussion and results in Section V. The conclusion of this work is in Section VI. Fig. 1: Octave of scale space [1] Fig. 2: Local maxima and minima of DOG [1] II. SIFT ALGORITHM SIFT algorithm contains the following stages:- 1) Scale-space extrema detection: In order to imple- ment the first stage, it was important to detect features at different scales. Therefore, one good choice was Gaussian kernel(G). An image (I ) will be convolved with that kernel as shown in Equation 1:- L(x, y, σ)= G(x, y, σ) ∗ I (x, y) (1) D(x, y, σ)=(G(x, y, k σ ) − G(x, y, σ)) ∗ I (x, y) = L(x, y, k σ ) − L(x, y, σ). (2) The Equation 2 represents difference of Gaussian (DOG), which is considered a decent approxima- tion to the scale-normalized Palladian of Gaussian σ 2 ▽ 2 G. After that, an octave of scale space was built. 16th international conference on Sciences and Techniques of Automatic control & computer engineering - STA'2015, Monastir, Tunisia, December 21-23, 2015 STA'2015-PID3628-IFR 978-1-4673-9234-1/15/$31.00 ©2015 IEEE 368