Instructions for Participants



Each participant is invited to submit its own algorithm in a Win32 console application. The performance will be evaluated by utilizing a very large data set of hand images captured with a normal flatbed scanner, the EPSON PERFECTION V300 PHOTO. The binary file should open all the .tif contained in a specific folder indicated by the user, and output a results.txt file with each line containing the information of the detected and considered characteristic points for each image file (each image always has 5 tips and 4 valleys), as in the following example:

!,0000.tif,T Xt1 Yt1,T Xt2 Yt2,T Xt3 Yt3,T Xt4 Yt4,T Xt5 Yt5,V Xv1 Yv1,V Xv2 Yv2,V Xv3 Yv3,V Xv4 Yv4,?
!,0001.tif,T Xt1 Yt1,T Xt2 Yt2,T Xt3 Yt3,T Xt4 Yt4,V Xv1 Yv1,V Xv2 Yv2,?

where (Xti,Yti) and (Xvi,Yvi) represent the pixels coordinates of the ith detected tip and valley, respectively.
The first line in the example reports a case where all tips and valleys were detected and considered in image 0000.tif, while the second line reports a case where two valleys and one tip were not considered by the competitor in image 0001.tif, even if they were detected.
(for a more detailed example, please refer to the training database annotation file)

The participants should briefly indicate how to run the submitted software.

The evaluation will be based on measures that will consider the quantity and type of the detected characteristic points as well as their distance to the ground truth, which will be manually labeled beforehand. For a ground truth characteristic point, (x0;y0;t0), if a detected and considered characteristic point (x;y;t), satisfies (t=t0) and sqrt((x-x0)^2+(y-y0)^2)<20, it is said to be truly detected and results in a zero valued penalty. Otherwise, it is called a miss and results in a one valued penalty. If it was not detected, or detected but not considered, the penalty is 0.7.
  • The detection rate is defined as the ratio of truly detected characteristic points to all ground truth characteristic points.
  • The miss rate defined as the ratio of the number of missed characteristic points to the number of all ground truth characteristic points.
The performance rank will be compiled and published on the competition website and will be based on the penalties sum. The competitor with the lowest penalties sum will be the winner. The average distance error of the correctly detected characteristic points to the ground truth will be used in the cases where ties occur.

The best algorithm will win the "Best Hand Characteristic Points Detection Competition" at ICIAR 2011. A Special Session will be devoted to present and discuss the experimental results of the best algorithms (up to five). The authors of the best algorithms will be invited to submit a paper to the conference, benefiting a 50% discount.

The goal of the competition is to compare different methodologies for software-based hand characteristic points detection within a data set composed by real hand images. The ambition of the competition is to become the reference event for academic and industrial researches.

It is expected a strong contribution to the state of the art of this particular subject, as well as an improvement in the reliability of this crucial step in hand-based biometric security systems.