Deep Learning for 3D Hand Biometric Systems

Staff - Faculty of Informatics

Date: 18 May 2020 / 13:30 - 15:30

You are cordially invited to attend the PhD Dissertation Defense of Jan Svoboda on Monday May 18th, 2020 at 13:30.
Please note that given the updated Covid-19 restrictions, the proposal review will be held entirely online.
You can join here (meeting ID: 849 3922 6215 / Password: 765129).

Hands are an indispensable part of human bodies used in our everyday life to express ourselves and manipulate the surrounding world. Moreover, a hand contains highly-unique characteristics that allow for distinguishing among different individuals. Though fingerprints are widely-known for this, the hand also has a unique geometric shape, palmprint, and vein structure. The shapes of a hand and its parts have been studied as biometric identifiers in the past, yielding state-of-the-art approaches in the field of Biometrics; however, these are limited in their practicality. This thesis aims at overcoming the limitations and improving the performance of hand-shape-based biometric systems, taking advantage of the latest developments in Deep Learning and 3D Sensing. In particular, such systems can be improved either by means of biometric fusion or by working directly with the hand shape. Biometric fusion allows for improvement in hand-shape-based biometric systems by simultaneously utilizing other modalities such as palmprints or fingerprints. To this end, we propose novel deep-learning based approaches to contact-free palmprint and fingerprint recognition. These improvements achieve state-of-the-art results on multiple standard benchmarks in both palmprint recognition and in processing latent fingerprint impressions. Additionally, current 3D-sensing technologies provide low-cost sensors that allow us to capture the surface of the hand as a 3D point cloud. This type of data is often noisy but can be dealt with by employing recent deep-learning architectures designed for point clouds. We base our solution on state-of-the-art geometric deep learning architectures, extending them with a novel clustering layer. We show how to train an optimal representation of a noisy 3D point cloud of a human hand purely from synthetic data. For evaluation, we collect a brand-new dataset of human hand videos in RGB-D, named NNHand RGB-D. Extensive evaluation of our approach on our dataset unveils the viable potential of low-precision, hand-shape-based biometric systems.

Dissertation Committee:
- Prof. Michael Bronstein, Università della Svizzera italiana, Switzerland (Research Advisor)
- Dr. Jonathan Masci, Università della Svizzera italiana, Switzerland (Research co-Advisor)
- Prof. Kai Hormann, Università della Svizzera italiana, Switzerland (Internal Member)
- Prof. Jürgen Schmidhuber, Università della Svizzera italiana, Switzerland (Internal Member)
- Dr. Andrew Fitzgibbon, Microsoft Research Cambrdige, UK (External Member)
- Prof. Mark Nixon, University of Southhampton, UK (External Member)