With supporting and pressing by industrial and government, evolving biometrics architectures, and developing mobile sensors, the contactless major finger knuckle (proximal interphalangeal joint) research and technology continue to mature rapidly, and many of the applications of finger knuckle have achieved higher level of accuracy performance on the contactless deployment. Especially, the completely contactless biometrics systems have been driven by the rapid development of mobile sensor, while the completely contactless development has 6 DOF. Using major finger knuckle to idenfity the human identity on the completely contactless deployment, the crease pattern of finger knuckle region is easily deformed while the finger is not fixed on any device. In this scenario, the matching performance of finger knuckle is poor, and the current methods cannnot get acceptable performance on the deformable finger knuckle dataset [1]. However, each subject of the deformabel finger knuckle dataset only offer 6 samples by bending finger with 6 different angles. Meanwhile, the number of samples of the current public finger knuckle dataset is too small so that hard to train a large model to get higher generazational ability, and the current SOTA finger knuckle model is shallow. Therefore, for improving the matching performance and enlarging the datasetthe of completely contactless finger knuckle dataset, we caputured the largest, first, and multipose deformable finger knuckle video dataset.
To simulate the deformation of finger knuckle in the completely contactless manner, each client needs to slowly bend their finger from 0 degrees to about 90 degrees and then back from 90 degrees to 0 degrees, repeating this motion 2 to 3 times on a about 10 seconds 4K video by iPhone 12 or Samsung. Although the background of the video is simple, it will not be fixed with only black or white background. Meanwhile, the illumination will also be changed with indoor with lights and outdoor sunlight. And each client client needs to offer their little, middle, and ring finger of left and right hand. Therefore, the captured video are named by L4 (ring finger of left hand), L3 (middle finger of left hand), L2 (index finger of left hand), R2 (index finger of right hand), R3 (middle finger of right hand), R4 (ring finger of right hand), and saved under the subject ID folder. In this kind of deployment, we captured 351 different subjects with 805,768 image frames, and most of subjects are from the China while some are from India and Europe.
Due to data privacy issues, we automatically removed the background and audio of the original video. And for speeding up the response of this webpage and reducing the memory occupation, we have converted the 4K
video to the 480P for a quick visualization. If you want to go through the database, please download the
source 4K video database. Because the source 4K video database can show more detail crease feature of the
major finger knuckle which is critical feature for increasing the matching performance. We show the finger
knuckle video sample of 6 different fingers of 12 different subjects on the below:
While the finger is bending as shown on the "Sample Videos", the crease feature of the finger knuckle is easily deformed resulting in poor performance by matching texture feature. The most reliable and robust of feature of deformed finger knuckle is the keypoint or interest point with local feature, and the keypoint is scale, view angle, and illumination invariant. Thefore, we use deep learning module to get the correspondence between image pairs. For a quick impression, we also visualize the correspondence as below:
The database is being made available for the researchers from July 2024 onwards. Interested researchers should follow following steps to acquire "The Hong Kong Polytechnic University Completely Contactless Multipose Finger Knuckle Video Database"
All the rights of the The The Hong Kong Polytechnic University Completely Contactless Multipose Finger Knuckle Video Database are reserved and commercial use/distribution of this database is strictly prohibited. All the technical reports and papers that report experimental results from this database should provide due acknowledgement and reference. The reference should appear as 'The Hong Kong Polytechnic University Completely Contactless Multipose Finger Knuckle Video Database'. Questions regarding this database can be directed to Ajay.Kumar@polyu.edu.hk.
[1] A. Kumar, 'Toward pose invariant and completely contactless finger knuckle recognition,' IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 1, no. 3, pp. 201-209, July 2019.
Back to
Databases
Page