ICDAR 2019
Competition on Signature Verification
based on
an On-line and Off-line Signature Dataset

Competition terms

Since the successful SigComp2009, the first signature verification competition on the ICDAR conference series, three further competitions have been organized in a row (SigComp2011, SigWIComp2013 and SigWIComp2015).

Hereby we announce the SigComp2019 signature verification and gender detection competition using on-line and off-line signatures.

DATASET

The SigComp2019 dataset consists of Hungarian signatures recorded by a Wacom Mobile Pro Studio 13 with Inking Pen.

Each signer contributed 8-16 genuine signatures in single sessions.

The on-line and off-line signatures were recorded simultaneously on sheets of A4 papers with the arrangement given in the SigComp2019_A4_training.pdf file. Each paper had 8 numbered rectangles with grey backgrounds around with the size of 25x100mm. The signatures were scanned with 600 DPI, RGB colour in TIFF format. The on-line and the scanned off-line signatures were separated one-by-one, thus one file contains exactly one signature. Each on-line signature was shifted. The bottom left corner of the black rectangle was used as origin.

The on-line signatures are stored in ASCII files in CSV format, values are separated by a semicolon (;). Each on-line data file has a header which comprises event type (EventType), timestamp (T), x and y coordinate (X and Y), pen-tip force aka pressure (P), altitude and azimuth (Altitude and Azimuth) and distance from the surface of the tablet (Distance).

EventType values:

  • 1. the pen is on the tablet
  • 3. the pen is lifted, but within the reading height (pressure is 0)

Metadata

Metadata was collected about the signers: signerID, age and biological gender is provided in a CSV file.

Training dataset

The training dataset (SigComp2019-on-offline-train) consist only genuine signatures (no forgeries) from 21 signers, 8 signatures by signer acquired by Inking Pen in both on-line and off-line format. In total: 21*8 = 168 signatures.

You may train your algorithm on publicly available datasets as well. Therefore, teams which can not access the training set of SigComp2019 can still participate in the competition.

Test dataset

The evaluation dataset (SigComp2019-on-offline-test) consist of genuine and forged signatures from another 50 signers (different signers than in the training dataset). Per signer:

  • 6 reference signatures
  • 10 genuine signatures
  • 8-12 skilled forgeries
In total: 300 references, 500 genuine and 400-600 skilled forgeries

Forged signatures were provided by the employees of Cursor Insight after training received from a handwriting expert who gave a recommendation on how to forge signatures properly. Every forger had a chance to see at least 5 reference signatures of the specimen and was allowed to practice the signatures as long as s/he wished to before the skilled forgeries were recorded. Every specimen was forged by at least two forgers.

Test on other datasets

We may test the submitted systems on other datasets as well.

Sharing the dataset

Due to data protection reasons, we are allowed to share directly the dataset only with the participants who sign a declaration which states that they handle the data according to the attached Privacy policy. Please read the policy and send the signed declaration to sigcomp2019@cursorinsight.com. We send in return the dataset that you requested.

According to the GDPR {General Data Protection Regulation}, the data sharing must meet sections 44-49 and 50 of the regulation. We can not send data out of the EU apart from these countries: Andorra, Argentina, Canada (only commercial organisations), Faroe Islands, Guernsey, Israel, Isle of Man, Jersey, New Zealand, Switzerland, Uruguay and USA (if the recipient belongs to the Privacy Shield). Any other enquiry will unfortunately be refused.

In accordance with the Competition Terms, Cursor Insight will not share the Dataset with Participants unless they provide evidence that sharing the Dataset with them complies with Sections 44-50 of GDPR relating to the transfer of personal data.

  • EU business participants may provide European Unique Identifier (EUID)
  • Non-EU business participants may provide business register extract and if necessary, other document stating their compliance with GDPR
  • All other participants (e.g. academia) may provide any documentation confirming their place of activity

Due to the strict regulation, we updated the competition terms at the "Training dataset" section, thus one may train their algorithm on publicly available datasets as well. This makes it possible to participate in the competition, evenif the GDPR does not make it possible to access the recent SigComp2019 training dataset.

Attachments

TASKS

Task 1/a and Task 1/b verification (on-line and off-line)

During each evaluation (each particular program run), the verifier will receive one questioned on-line (in Task 1/a) or off-line (in Task 1/b) signature from the evaluation dataset (that can be a genuine or a forged signature) and 3-6 reference signatures. The verifier has to report a comparison score (degree of similarity) and an evidential value of that score.

The evidential score was applied in the previous competitions and it is expressed as the ratio of the probabilities of finding that comparison score when the questioned signature belongs to the acclaimed author and when the questioned signature belongs to someone else.

Task 2/a and Task 2/b: gender detection (on-line and off-line)

During each evaluation (each particular program run), the classifier will receive one genuine on-line signature from the evaluation dataset. The verifier has to report a decision about whether the signature was written by a female or a male signer.

EVALUATION

PLANNED SCHEDULE

  • The release of the training dataset: April 2019
  • System submission: 30th April 2019
  • Conference: 20th-25th September 2019

Organisers

  • Gergely Hancz√°r PhD, Cursor Insight Ltd
  • Erika Griechisch PhD, Cursor Insight Ltd

Consultants

  • Muhammad Imran Malik, PhD
  • Sheraz Ahmed, PhD

Email us for updates: sigcomp2019@cursorinsight.com