嶄猟井


Main

Research

People

Demos

Publications

Contact us


 

 

Last Update :
2011-1-20

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

FAST COMMON VISUAL PATTERN DETECTION VIA RADIATE GEOMETRIC MODEL

 

ALGORITHM

Radiate Geometric Structure (RGS)

The two forms of

Radiate Geometric Structure (RGS)

are equivalent to each other.

The geometrical form of a Radiate Geometric Structure The mathematical form of a Radiate Geometric Structure

Radiate Geometric Model (RGM)

Radiate Geometric Model (RGM)

A RGM consists of a set of RGS which is obtained by considering each pair of matched points as referrence center.

RGS_1
a
RGS_2
a
...
...
RGS_i
a
...
...
RGS_N
a

Transform into Graph

arrow

One dense sub-graph correspond to one common visual pattern. By detection all dense sub graphs, we can complete the task of CVP detection.

DATA-SET

(To download the full pakage of original data, click here ~)

The data set we used is a subset of [1]. There are 10 kinds of images organized in 10 different folders in the data set of [1] .

Our data set is obtained by the following steps:

1.We choose 4 folders from [1] such as the folder of 'Google', 'KFC', 'Monalisa' and 'Starbark'.

2.For each one of the 4 folders, we randomly choose about 10 images to form the four sets of original data.

3.The final data set for experiment is made up of the original data by combining every two images in the same folder to make a couple.

Note: Because the local feature matching method we used is not symetrical, for a pair of images (img1, img2), we do the experiment on both (img1,img2) and (img2,img1).

[1] http://www.jdl.ac.cn/en/project/mrhomepage/IPDID.htm

Several images in the original data sets are shown as follows:

Google
......
KFC ......
Monalisa ......
Starbark ......

 

EXPERIMENTAL RESULTS

To download the full pakage of experimental results, click here ~

The manually labeled results are also attached to the full package so that it would be easy for others to compare with our method.

Questions are welcomed, please contact lychu@jdl.ac.cn.

Google

a b

KFC

D F E

Monalisa

G H

Starbark

V Y U

 

COMPARING RESULTS

The comparing results consists of two factors: computing time and detection accuracy.

The method to evaluate computing time:

1.Match the local feature points between the given pair of partial duplicate images.

(Note: we use the same feature matching method for both the two compared algorithms and the computing time of feature mathing is not considered)

2.The computing time of CVP detection is recorded by the programs.

The method to evaluate detection accuracy:

1. Manually label the number of validly matched feature point pairs.

2. Manually label the number of validly detected object pairs (CVPs).

3. Calculate the numerical evaluation of the detection accuracy of both the two compared algorithms.

The figure on the left is the comparing result between our method and the method of [1]. each point corresponds to a pair of near duplicate images.The x-coordinate shows the comparing result of computing time and the y-coordinate shows the comparing result of detection accuracy.

The table on the top is the average time ratio and average accuracy ratio.

The following results can be easily figured out from the Figure and Table:

1.Our method is at least 40 times faster than the method of [2];

2.Our method has achieved a better performance of detection accuracy;

[2] Hairong Liu and Shuicheng Yan. Common Visual Pattern Discovery via Spatially Coherent Correspondences. CVPR 2010.

 

COPYRIGHT

Internet Partial-Duplicate Image Dataset is for academic research only. JDL is the copyright holder of all the images included in this dataset. If you use this dataset, please cite the following paper:

Lingyang Chu, Shuqiang Jiang, Qingming Huang, "Fast Common Visual Pattern Detection via Radiate Geometric Model", IEEE International Conference on Image Processing, January, ICIP 2011. (submitted)