Download the license agreement (next to the database description), fill it out, sign it by a faculty member, and send it by email to gpds@gi.ulpgc.es, as follows:
Subject: TOOLBOX/DATABASE download
Body: your name, e-mail, phone number, organisation, postal address, the database you require, the purpose for which you will use it, and the date you sent the signed license agreement. Attach a PDF of the signed agreement to the email.
Once we receive the license agreement, you will get an email with instructions to download the database. After downloading, please notify us by email that the transaction was completed successfully.
3D-iDeMLog is a representation of 3D handwriting based on the lognormality principle, which represents the peak pattern of the velocity in well-learned, rapid human movement as an overlapped sum of lognormals. It extends the 3D Sigma-Lognormal by, firstly, detecting and splitting multi-stroke speed peaks into single strokes; secondly, approximating the single strokes with single lognormals; thirdly, modelling multi-lognormal strokes with several lognormals having a common time of occurrence; and finally, generalizing the planar trajectories between virtual target points.
Miguel A. Ferrer, Moises Diaz , Jose Juan Quintana, Cristina Carmona-Duarte, Réjean Plamondon (2023), "A Multi-Lognormal Analysis of 3D Handwriting", Knowledge-Based Systems. doi: To appear
Miguel A. Ferrer, Moises Diaz , Jose Juan Quintana, Cristina Carmona-Duarte, Réjean Plamondon (2023), "Synthesis of 3D On-Air Signatures with the Sigma-Lognormal Model", Knowledge-Based Systems. Vol 265, pp. 110365. doi: 10.1016/j.knosys.2023.110365
Miguel A. Ferrer, Moises Diaz, Cristina Carmona-Duarte, Jose Juan Quintana and Réjean Plamondon. "iDeLog3D: Sigma-Lognormal Analysis of 3D Human Movements". In: Carmona-Duarte, C., Diaz, M., Ferrer, M.A., Morales, A. (eds) Intertwining Graphonomics with Human Movements. IGS 2022. Lecture Notes in Computer Science, vol 13424. pp 189–202. Springer, Cham. doi: https://doi.org/10.1007/978-3-031-19745-1_14
3D-iDeMLog is developed in Matlab and its source code can be freely downloaded for research purposes. Once you send a signed copy of the license agreement below to gpds@gi.ulpgc.es, we will email you the source code.
iDeLog 2D is a novel framework to extract the Sigma-Lognormal parameters for 2D spatiotemporal signals like handwriting. Specifically, iDeLog consists of two steps. The first one, influenced by the motor equivalence model, separately derives an initial action plan defined by a set of virtual points and angles from the trajectory and a sequence of lognormals from the velocity. In the second step, based on a hypothetical visual feedback compatible with an open-loop motor control, the virtual target points of the action plan are iteratively moved to improve the matching between the observed and reconstructed trajectory and velocity.
Ferrer, M. A., Diaz, M., Carmona-Duarte, C., and Plamondon, R. (2018). iDeLog: iterative dual spatial and kinematic extraction of sigma-lognormal parameters. IEEE Transactions on Pattern Analysis and Machine Intelligence. 42(1), 114-125. doi: 10.1109/TPAMI.2018.2879312
Ferrer, M. A., Diaz, M., Quintana, J. J., Carmona-Duarte, C., and Plamondon, R. (2023). Extending the kinematic theory of rapid movements with new primitives. Pattern Recognition Letters, 167, 181-188. doi: 10.1016/j.patrec.2023.02.021
iDeLog2D is developed in Matlab and its source code can be freely downloaded for research purposes. Once you send a signed copy of the license agreement below to gpds@gi.ulpgc.es, we will email you the source code.
Novel Anthropomorphic Features for On-line Signatures
A novel feature space for on-line signatures is available for research purposes. The features characterise the movement of the shoulder, elbow and wrist joints when signing, through a virtual skeletal arm (VSA) model that simulates the architecture of a real arm and forearm. They are divided into Position-based and Angle-based anthropomorphic features.
The procedure for obtaining these features is detailed in:
M. Diaz, M.A. Ferrer and J.J. Quintana (2018). Anthropomorphic Features for On-line Signatures. IEEE Transactions on Pattern Analysis and Machine Intelligence. doi: 10.1109/TPAMI.2018.2869163
M. Diaz, M.A. Ferrer and J.J. Quintana. "Robotic Arm Motion for Verifying Signatures". 16th International Conference on Frontiers in Handwriting Recognition (ICFHR), Niagara Falls, USA, 5–8 August 2018, pp. 157–162. doi: 10.1109/ICFHR-2018.2018.00036
To obtain the anthropomorphic features, run the following in Matlab:
[Q, pe, pw, pf] = pos2angL2L4(x, y, z, az, in, L2, L4);
The inputs x, y, z correspond to the spatial trajectory of the signature in mm. L2, L4 denote the humerus and radius/ulna lengths. The output Q is a vector with the six angle-based features. pe, pw, pf denote the 3D position of the elbow, wrist and fingers, respectively.
The function pos2angL2L4 is developed in Matlab and its source code can be freely downloaded for research purposes. Once you send a signed copy of the license agreement below to gpds@gi.ulpgc.es, we will email you the source code.
A synthetic duplicator engine for static signatures is available. It is developed in Matlab and distributed in *.p format.
The duplicator is based on a set of nonlinear and linear transformations that simulate the human spatial cognitive map and the motor-system intra-personal variability during the signing process.
The procedure is detailed in:
M. Diaz, M.A. Ferrer, G. Eskander, R. Sabourin (2016). "Generation of Duplicated Off-line Signature Images for Verification Systems". IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 5, pp. 951–964. doi: 10.1109/TPAMI.2016.2560810
M. Diaz, M.A. Ferrer and R. Sabourin (2016). "Approaching the Intra-Class Variability in Multi-Script Static Signature Evaluation". 23rd International Conference on Pattern Recognition (ICPR), Cancun, Mexico.
To use the duplicator, run the following in Matlab:
A Hidden Markov Model (HMM) toolbox for the Matlab environment is available. It implements conventional continuous and discrete HMM techniques for both training and testing, supports grouping components for the feature vector, and provides multi-labelling for the discrete HMM. The toolbox includes procedures suitable for classical HMM applications such as pattern recognition, speech recognition and DNA sequence analysis.
The toolbox is distributed as binary (dll files) and source code. For wider promotion, we ask users to reference:
Sébastien David, Miguel A. Ferrer, Carlos M. Travieso, Jesús B. Alonso. "gpdsHMM: A hidden Markov model toolbox in the Matlab environment". International Conference on Complex System Intelligence and Modern Technology Application, Cherbourg, Normandy, France, 19–22 September 2004.
Dual off-line and on-line signature databases of Bengali and Devanagari signatures. They contain data from 100 synthetic individuals, with 24 genuine signatures per individual. All static signatures were generated with different modelled pens. The synthetic users were generated following the procedure described at:
M.A. Ferrer, S. Chanda, M. Diaz, C.Kr. Banerjee, A. Majumdar, C. Carmona-Duarte, P. Acharya and U. Pal (2018). "Static and Dynamic Synthesis of Bengali and Devanagari Signatures". IEEE Transactions on Cybernetics, vol. 48, no. 10, pp. 2896–2907. doi: 10.1109/TCYB.2017.2751740
Moises Diaz, Sukalpa Chanda, Miguel Ferrer, Chayan Kr. Banerjee, Anirban Majumdar, Cristina Carmona-Duarte, Parikshit Acharya and Umapada Pal. "Multiple Generation of Bengali Static Signatures". 15th ICFHR, Shenzhen, China, 23–26 October 2016, pp. 42–47. doi: 10.1109/ICFHR.2016.0021
Static signatures are in JPG format at an equivalent resolution of 600 dpi; genuine files are named xxx/c-xxx-yy.jpg (xxx = signer number, yy = repetition). Dynamic signatures are MAT files containing the x/y coordinates sampled at 100 Hz plus the pressure sequence p; genuine files are named xxx/c-xxx-yy.mat.
For performance reference, the table below is an excerpt of the reference above:
EER (%) by training-set size, classifier and script
Training
Classifier
Bengali
Devanagari
Real 1
Real 2
Synthetic
Real 1
Real 2
Synthetic
2
HMM
6.03
5.54
6.24
5.32
6.71
5.73
SVM
4.32
0.84
2.73
2.85
1.37
2.03
DTW
NA
0.41
1.66
NA
0.41
1.02
Man
NA
8.19
12.6
NA
9.77
12.2
5
HMM
4.08
3.18
3.06
3.37
4.43
2.68
SVM
1.97
0.23
0.67
1.56
0.56
0.47
DTW
NA
0.27
0.47
NA
0.31
0.49
Man
NA
3.41
6.50
NA
4.16
5.79
8
HMM
3.17
2.46
2.38
2.77
3.58
1.75
SVM
1.32
0.13
0.34
1.37
0.38
0.24
DTW
NA
0.33
0.25
NA
0.34
0.36
Man
NA
2.7
5.44
NA
3.59
4.7
10
HMM
2.5
2.34
1.71
2.77
3.22
1.34
SVM
1.12
0.13
0.25
1.33
0.31
0.14
DTW
NA
0.28
0.23
NA
0.29
0.38
Man
NA
2.57
5.11
NA
3.11
4.41
NA stands for "Not Apply", as dataset 1 does not contain on-line signatures.
Dual off-line and on-line signature database. It contains data from 10,000 synthetic individuals: 24 genuine signatures plus 30 forgeries per individual. All static signatures were generated with different modelled pens. The synthetic users were generated following the procedure described at:
M.A. Ferrer, M. Diaz, C. Carmona-Duarte, A. Morales (2016). "A Behavioral Handwriting Model for Static and Dynamic Signature Synthesis". IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1041–1053. doi: 10.1109/TPAMI.2016.2582167
Static signatures are PNG files at an equivalent resolution of 600 dpi. Genuine files: xxx\c-xxx-yy.png; forgery files: xxx\cf-xxx-yy.png. Dynamic signatures are MAT files (x/y at 100 Hz plus pressure p): genuine xxx\c-xxx-yy.mat, forgeries xxx\cf-xxx-yy.mat.
For performance reference, verifiers were trained following the established protocol of [2]: 5 randomly selected genuine signatures for training, remaining genuine signatures for the false-rejection-rate test. The false-acceptance rate for the random-impostor experiment used genuine test samples from all other users; the deliberate-forgeries experiment used each signer's forgery samples.
Four classifiers were compared — two off-line, two on-line: HMM with geometrical features [1]; SVM with Local Binary Pattern features [2]; on-line Dynamic Time Warping (DTW) with Euclidean distance over [x, y, p, dx, dy, dp, ddx, ddy, ddp] [3]; and a Manhattan-distance histogram-based classifier [4].
M.A. Ferrer, J.B. Alonso and C.M. Travieso, "Offline Geometric Parameters for Automatic Signature Verification using Fixed-Point Arithmetic", IEEE TPAMI, vol. 27, no. 6, pp. 993–997, June 2005.
M.A. Ferrer, F. Vargas, A. Morales and A. Ordoñez, "Robustness of Off-line Signature Verification Based on Gray Level Features", IEEE TIFS, vol. 7, no. 3, pp. 966–977, June 2012.
A. Fischer, M. Diaz-Cabrera, R. Plamondon, M.A. Ferrer, "Robust Score Normalization for DTW-Based On-Line Signature Verification", 13th ICDAR, Tunis, Tunisia, August 2015.
N. Sae-Bae and N. Memon, "Online Signature Verification on Mobile Devices", IEEE TIFS, vol. 9, no. 6, pp. 933–947, June 2014.
Results in terms of EER (%):
EER (%) — random impostors vs. deliberate forgeries
Off-line signature database. It contains data from 4,000 synthetic individuals: 24 genuine signatures plus 30 forgeries per individual, all generated with different modelled pens. This database replaces previous signature databases. The synthetic users were generated following the procedure described at:
Miguel A. Ferrer, Moisés Diaz-Cabrera, Aythami Morales. "Static Signature Synthesis: A Neuromotor Inspired Approach for Biometrics". IEEE Transactions on Pattern Analysis and Machine Intelligence, ISSN 0162-8828, vol. 37, no. 3, pp. 667–680, March 2015.
Signatures are in JPG format at an equivalent resolution of 600 dpi. Genuine files: xxx\c-xxx-yy.png; forgery files: xxx\cf-xxx-yy.png.
For performance reference, results on previous public databases using the same verifier as in:
Miguel A. Ferrer, Francisco Vargas, Aythami Morales, Aaron Ordoñez. "Robustness of Off-line Signature Verification based on Gray Level Features". IEEE Transactions on Information Forensics and Security, vol. 7, no. 3, pp. 966–977, June 2012.
Training with 10 genuine samples — public databases
Database
Users
Random forgeries
Simulated forgeries
MCYT
75
0.35%
11.54%
GPDS960
75
0.37%
13.78%
GPDS960
150
0.44%
15.90%
GPDS960
881
0.88%
23.42%
Training with 10 genuine samples — GPDSsyntheticSignature database
The database consists of 10 different acquisitions of 150 people, taken with a desk scanner. The 1,500 images were taken from the users' right hand. Users could place the palm freely over the scanning surface — no pegs, templates, or other constraining capture aids were used. The hand contour with landmarks (valleys and fingertips) and the segmented palms are also provided; they were obtained automatically, without supervision, as described in:
Miguel A. Ferrer, Aythami Morales, Carlos M. Travieso, Jesús B. Alonso. "Low Cost Multimodal Biometric Identification System based on Hand Geometry, Palm and Finger Textures". 41st Annual IEEE International Carnahan Conference on Security Technology, ISBN 1-4244-1129-7, pp. 52–58, Ottawa, Canada, 8–11 October 2007.
Images are JPG, 256 gray levels, 120 dpi. Hand files: xxx\manoxxx_yy.jpg; palm files: xxx\palmaxxx_yy.jpg. Hand contour and landmarks are in the Matlab2007 file metadata.mat; usage is shown in ReadDatabase.m.
This hands database consists of 10×3 acquisitions (visible, 850 nm and 1470 nm bands) from 100 people — 3,000 images of the users' right hand. Most users are 23–40 years old, with roughly half male. Users could place the hand freely over the plate, with no pegs, templates, or other constraining capture aids. The cameras acquire the hand dorsum image.
The 1470 nm band is acquired with a XENICS XEVA 1.7-320 camera (InGaAs sensor, sensitive 900–1700 nm) with a 1470 nm band-pass filter (250 nm bandwidth). The visible band is acquired with a Logitech QuickCam E2500 colour webcam (640×480). The NIR band is acquired with the same webcam fitted with a high-pass filter with an 850 nm cutoff. The procedure is described in:
Miguel A. Ferrer, Aythami Morales. "Hand Shape Biometrics combining the Visible and Short Wave Infrared Bands". IEEE Transactions on Information Forensics & Security (accepted).
Images are BMP, as given by the cameras. Files are named yyy\xxx_yyy_rr.bmp where xxx is the band (vis, 850 or 1450), yyy the user number, and rr the repetition.
This contactless hands database consists of 10×2 acquisitions (visible and 850 nm) from 100 people — 2,000 images of the users' right hand. Most users are 23–40 years old, with roughly half male. The user holds a hand over the camera and touchlessly adjusts its position to overlap with an on-screen hand mask; once overlap exceeds 70%, the device automatically captures both the IR and visible images. Details in:
Miguel A. Ferrer, Francisco Vargas, Aythami Morales. "BiSpectral Contactless Hand Based Biometric System". 2nd National Conference on Telecommunications (CONATEL 2011), Arequipa, Peru, 17–20 May 2011.
The acquisition device uses two inexpensive standard webcams capturing simultaneously: an "IR" webcam (750–1000 nm, made by removing the IR-blocking element and adding Kodak filters No. 87 FS4-518 and No. 87c FS4-519, no transmittance below 750 nm) and a "V" webcam (visible range, 400–700 nm). Infrared illumination comes from 24 GaAs IR-emitting diodes (CQY 99, peak wavelength 925 nm, 40 nm bandwidth) arranged in an inverted-U shape around the cameras; the open side of the U coincides with the wrists. The IR webcam focus is adjusted manually on first use.
Images are BMP, as given by the webcams. Visible band: xxx\visible_xxx_yy.bmp; infrared band: xxx\Infraro_xxx_yy.bmp; segmented palm: xxx\palma_xxx_yy.jpg; visible-image contour: xxx\Icontvisible_xxx_yy. Usage is shown in ReadDatabase.m.
The GPDShandsSWIRhyperspectral database consists of 10 samples from 154 people. Each sample comprises 350 hyperspectral images, for a total of 154 × 10 × 350 = 539,000 images across 256 bands between 900 nm and 1600 nm. Users were allowed to wear wristwatches or wristlets. Ages ranged from 18 to 60 (86% between 18 and 30); approximately 70% are students/teachers and 30% are administration/cleaning staff from the university, and 64% of users are male. For the first 24 users, hyperspectral images were acquired from each hand side (palm and dorsum).
More information:
Miguel A. Ferrer, Aythami Morales and Alba Díaz. "An Approach to SWIR Hyperspectral Hand Biometrics". Information Sciences. doi: 10.1016/j.ins.2013.10.011
The acquisition device is a Xenics® Xeva-1.7-320 camera (InGaAs detector, sensitive 900–1700 nm, 256 gray levels, 320×256 resolution), used with a SPECIM® ImSpector N17E spectrograph (f/2.0), turning the SWIR camera into a line spectral imager. The spectrographic images span 320 pixels in x and 256 wavelength bands; at f/2.0 with a 41 cm lens-to-plate distance, the x-dimension covers 21 cm (≈38 dpi horizontal resolution). A rotating mirror scanner attached to the spectrograph objective adds the y-axis: as the mirror rotates through 40°, the y-dimension covers 29 cm.
Acquisition procedure: the user places the hand dorsum on the plate, palm up. Once still, the mirror rotates from −20° to +20°, with the camera capturing about 12 frames per second over ~30 seconds, yielding 350 images in the y-axis (≈30 dpi vertical resolution); a reconstruction stage assembles a coherent hyperspectral cube.
The XEVA 1.7-320 had 99% pixel operability; the 298 erroneous pixels were interpolated as the mean of the pixels immediately above and below (not all surrounding pixels, to avoid row-contiguous errors). Illumination comes from two tungsten-filament bulbs (400–1600 nm spectrum, mostly NIR energy), placed symmetrically on each side of the hand to avoid shadows or uneven lighting.
10 acquisitions of 102 people, taken in two sessions one week apart (five samples each). 1,020 images of the users' right hand. The near-infrared hand-dorsum capture system uses two 64-LED arrays at 850 nm, a CCD Gigabit-Ethernet PULNIX TM3275 camera with a high-pass IR filter (750 nm cutoff), and a handle with two positional-reference pegs, as described in:
Miguel A. Ferrer, Aythami Morales, Lourdes Ortega. "Infrared Hand Dorsum Images for Identification". Electronics Letters, vol. 45, no. 6, pp. 306–308, ISSN 0013-5194, 12 March 2009.
Images are BMP, named xxx\mano-xxx-yyy.bmp. A readdatabase.m file is provided.
GPDS100VeinsCMOScylindrical database
10 acquisitions of 103 people in two sessions one week apart (five samples each); 1,030 images of the right hand. Capture system: two 64-LED arrays at 850 nm and a CMOS webcam with a high-pass IR filter (750 nm cutoff), with a cylindrical handle and two positional pegs. The users of the CMOScylindrical and CMOSergonomic databases are the same.
Images are BMP, named xxx\venas_xxx_yy.bmp. A readdatabase.m file is provided.
GPDS100VeinsCMOSergonomic database
Same acquisition protocol as above (103 people, two sessions, 1,030 images, two 64-LED arrays at 850 nm, CMOS webcam with 750 nm high-pass IR filter), but using an ergonomic handle that fixes the hand position comfortably. The users are the same as in the CMOScylindrical database.
Images are BMP, named xxx\venas_xxx_yy.bmp. A readdatabase.m file is provided.
The database consists of 10 different acquisitions of the face of 51 people, captured with a CCD camera to highlight the lips, as described in:
Carlos M. Travieso, J. Zhang, P. Miller, Jesús B. Alonso, Miguel A. Ferrer. "Bimodal Biometric Verification Based on Face and Lips". Neurocomputing, ISSN 0925-2312, vol. 74, pp. 2407–2410, 2 June 2011.
Pictures are JPG, named xxx\labio-xx-yy.jpg. A Matlab reader, ReadDatabase.m, is provided.
gpds SIGNATURE CORPORA
GPDS960signature database
This database is no longer available due to the General Data Protection Regulation (EU) 2016/679 (GDPR). Instead, you can apply for the off-line GPDSsyntheticSignature database or the GPDSsyntheticOnLineOffLineSignature database. We apologise for the inconvenience.
For historical reference, the database was an off-line signature collection of 960 individuals: 24 genuine signatures per individual, collected in a single-day session, plus 30 forgeries each. Forgeries were produced from the static image of the genuine signature; each forger could practise as long as desired and imitated 3 signatures from 5 (randomly chosen) signers in a single session, so each genuine signature had 30 skilled forgeries made by 10 forgers from 10 different genuine specimens.
Images were BMP, black and white, 300 dpi: genuine xxx\c-xxx-yy.bmp, forgeries xxx\cf-xxx-yy.bmp. As the scanned background contrasted well with the darker strokes, images were binarised with a fixed threshold, and stray "hair" artefacts from the strokes were removed, per [1] below. A ReadDatabase.m file was provided.
This competition was held during the 12th ICFHR (Kolkata, India, 2010) using a subset of the GPDS960Signature database.
The training subset used 4 genuine signatures from individuals 301–700 of GPDS960signature, named Trainingset\xxx\c-xxx-yy.bmp (xxx = signer id 301–700, yy = repetition 01–04). A Matlab script to read and display the training images is provided: ReadTrainingSignatures.m.
For testing, 30,000 questioned signature images were used, drawn from GPDS960signature signers 301–700 (originals and simulated forgeries) plus random signatures from users 701–960. Test files are named c-xxxxx-yyy.bmp (xxxxx = file number 00001–30000, yyy = claimed identity 301–700). A Matlab script to read and display the test images is provided: ReadTestSigantures.m.
To evaluate a verifier against this dataset, a Matlab program EvaluateASV is provided; it requires asv.m and the file 4nSigCompSignatureIdentification.mat (containing the matrix sign). The expected function signature is:
function decision = asv(signature, id)
where signature is the image to verify and id the claimed identity; decision should be 1 if accepted as genuine, 0 if considered a forgery. Matrix conventions: for test file c-xxxx-yyy.bmp, sign(xxxx,1) is the true signer, sign(xxxx,2) the repetition, sign(xxxx,3) the claimed identity; sign(xxxx,4) equals 0 for a genuine signature, 1 for a simulated forgery, and 2 for a random forgery.
For further information, see [1] below.
[1] M. Blumenstein, Miguel A. Ferrer, J.F. Vargas. "The 4NSigComp2010 Off-line Signature Verification Competition: Scenario 2". 12th ICFHR, ISBN 978-0-7695-4221-8, pp. 721–726, Kolkata, India, 16–18 November 2010.
GPDS960GRAYsignature database
This database contains a gray-level version of the genuine signatures and imitations from the GPDS960Signature database. Following a relocation, signatures of 79 users and 143 imitations of the remaining signers were unfortunately lost. The database therefore consists of 881 users, 21,144 genuine signatures and 26,317 imitations (47,485 signatures in total); the lost users/imitations are listed in ReadDatabase.m.
Signatures are PNG, scanned at 600 dpi: genuine xxx\c-xxx-yy.png, forgeries xxx\cf-xxx-yy.png. As the sheets were rescanned at 600 dpi for this version, segmentation errors differ from those in the original GPDS960Signature database. This database was used in:
Miguel A. Ferrer, Francisco Vargas, Aythami Morales, Aaron Ordoñez. "Robustness of Off-line Signature Verification based on Gray Level Features". IEEE Transactions on Information Forensics and Security, vol. 7, no. 3, pp. 966–977, June 2012.
Checks database
Contains 20 images — 12 bank checks and 8 invoices — with varying background complexity. Used in the Ferrer et al. (2012) paper above; please reference it if you use this database.
The paper blended the MCYT database (atvs.ii.uam.es) and the GPDS960GRAYSignatures database with the check images, using the "multiply" blend mode, to create a synthetic database with distorted gray levels (each overlaid stroke produces a new, darker gray level).
The following Matlab files are provided:
ReadCheckDatabase.m — reads the check database, divided into the three gray-level distortions used in the paper.
BlendSignaturewithCheck.m — blends a signature with a given check.
LBP.m, LDP.m, LDeriv.m — texture-feature extractors (Local Binary Patterns, Local Directional Pattern, Local Derivative Pattern) used in the paper.
Acquired under laboratory conditions, this single-session database contains 380 latent palmprints from 100 different palms of 51 donors (28 male, 23 female), aged 4 to 81, with various occupations (manual workers, office workers, students). Each donor contributed two impressions (right and left hands) plus multiple latent prints simulating realistic poses — opening a door, pushing a chair, grasping a knife, leaning on a table, carrying objects of different weights, among others. For details, see:
Aythami Morales, Miguel Angel Medina-Perez, Miguel A. Ferrer, Milton Garcia-Borroto, Leopoldo Altamirano Robles. "LPIDB v1.0 — Latent Palmprint Identification Database". International Joint Conference on Biometrics, Florida, 2014.