MEASUREMENT SCIENCE REVIEW, Volume 13, No. 3, 2013
122
A Comparative Study of SIFT and its Variants
Jian Wu
1
, Zhiming Cui
1
, Victor S. Sheng
2
, Pengpeng Zhao
1
, Dongliang Su
1
, Shengrong Gong
1
1
The Institute of Intelligent Information Processing and Application, Soochow University, Suzhou 215006, China,
jianwu@suda.edu.cn
2
Department of Computer Science, University of Central Arkansas, Conway 72035, USA
SIFT is an image local feature description algorithm based on scale-space. Due to its strong matching ability, SIFT has many
applications in different fields, such as image retrieval, image stitching, and machine vision. After SIFT was proposed, researchers
have never stopped tuning it. The improved algorithms that have drawn a lot of attention are PCA-SIFT, GSIFT, CSIFT, SURF
and ASIFT. In this paper, we first systematically analyze SIFT and its variants. Then, we evaluate their performance in different
situations: scale change, rotation change, blur change, illumination change, and affine change. The experimental results show that
each has its own advantages. SIFT and CSIFT perform the best under scale and rotation change. CSIFT improves SIFT under
blur change and affine change, but not illumination change. GSIFT performs the best under blur change and illumination change.
ASIFT performs the best under affine change. PCA-SIFT is always the second in different situations. SURF performs the worst in
different situations, but runs the fastest.
Keywords: Image matching, local feature, SIFT, PCA-SIFT, GSIFT, CSIFT, SURF, ASIFT
1.
INTRODUCTION
MAGE MATCHING is an important research direction in
computer vision and image processing. It is also a
necessary precondition to solve many practical problems.
Many researchers are dedicated to improving the
performance of image matching techniques, and have
proposed a variety of algorithms [1]. The image matching
algorithms can be divided into two categories: global
feature-based matching algorithms and local feature-based
matching algorithms [2]. Comparing with global feature-
based matching algorithms, local feature-based matching
algorithms are more stable. They have been applied
successfully in many real-world applications, such as object
recognition, texture recognition, image retrieval, robot
localization, video data mining, building panoramas, and
object category recognition [3]-[5].
Local feature-based matching algorithms include two
stages: interest point detection and description. Good local
features should have the following proper characteristics.
Feature detection has a high repeatability rate and high
speed. Feature description has a low feature dimension,
which is easy to achieve quick matching and robustness to
illumination, rotation, and viewpoint change. David G.
Lowe proposed a local feature description algorithm SIFT
(Scale-invariant Feature Transform) [6], [7] based on the
analysis of existing invariance-based feature detection
methods at that time. SIFT has good stability and invariance.
It detects local keypoints, which contain a large amount of
information. Because of its unique advantages, it has
become a popular research topic. Many researchers
constantly work hard to improve it.
Tuytelaars and Mikolajczyk presented a decent overview
on most widely used local invariant feature detectors [8].
This survey article consists of two parts. After reviewing
local invariant feature detectors, it distinguishes among
corner detectors, blob detectors, and region detectors. Juan
and Gwun [9] summarized the three robust feature detection
methods: SIFT, PCA-SIFT, and SURF. The performance of
the robust feature detection methods is compared for scale
change, rotation change, blur change, illumination change,
and affine transformations. Younes et al. [10] discussed
three implementations of the SIFT algorithm, i.e., Lowe’s,
Hess’s, and theirs, respecting the parameterization suggested
by Lowe in 2004.
This paper makes a more in-depth analysis and
comparisons on the SIFT algorithm and its most concerned
five variants. In order to find out their advantages and
disadvantages, we conduct experiments to evaluate their
performance in different situations: scale change, rotation
change, blur change, illumination change, and affine change.
Their performance is evaluated in terms of a popular
measure: matching correct rate. We also further investigate
their time consumption. At the end of this paper, we discuss
their advantages and disadvantages, and make conclusions
on this study.
2.
RELATED WORK
Since the SIFT algorithm was formally proposed,
researchers have never stopped improving it. According to
the statistics of references in Google Scholar, the article [7],
which published the SIFT algorithm, has more than 12,000
references. Among its variants, the numbers of references of
PCA-SIFT [11], GSIFT [12], CSIFT [13], SURF [14] and
ASIFT [15] are relatively high. Thus, these algorithms are
selected and investigated in this paper.
The procedure of SIFT mainly includes three steps:
keypoint detection, descriptor establishing, and image
feature matching. Researchers improve the performance of
SIFT by adjusting these steps. Most of them just adjust one
of the three steps. Detailed discussions are as follows.
In the phase of descriptor establishing, SIFT uses a 128-
dimensional vector to describe each keypoint. This high
dimension makes the following step of SIFT (image feature
matching) slow. In order to reduce the dimensionality of
describing each keypoint, Y. Ke [11] uses the Principal
Component Analysis method to replace the histogram
I
10.2478/ms
-2013-0021
Brought to you by | Soochow University Library
Authenticated
Download Date | 1/12/16 7:07 AM