Bilkent University
Department of Computer Engineering
M.S.THESIS PRESENTATION
Joint Modeling of Local and Global Facial Dynamics for Hand Strength Estimation in Competitive Poker
Murat Şahin
Master Student
(Supervisor: Assoc.Prof.Hamdi Dibeklioğlu )
Computer Engineering Department
Bilkent University
Abstract: Facial signals play an important role in understanding an individual's internal state, particularly when they hold information that others seek to infer. Studying such signals in unscripted, real-life situations with meaningful stakes is challenging, as suitable video data is scarce and behavioral labels often rely on subjective judgment. We explore this problem through professional poker, using publicly available competitive livestreams where players act naturally on camera while objective labels can be derived from the game state. We introduce PokerFace, a dataset constructed by automatically processing hundreds of hours of broadcasts into over a thousand short facial clips labeled by hand strength from each player’s perspective. We propose a parameter-efficient architecture built on a frozen self-supervised backbone, combining global and local streams with lightweight cross-attention for temporal modeling. Together, the two streams preserve broad temporal context while capturing subtle facial expressions that might otherwise be missed. Under strict leave-one-subject-out evaluation on our collected dataset, our approach achieves promising results and remains competitive on two deception detection datasets, DOLOS and Box of Lies. The ablation studies support this design, showing that combining both streams yields the strongest overall performance. Our study shows that professional poker provides a valuable setting for studying facial dynamics and demonstrates how real-world competitive environments can help improve our understanding of human facial behavior. Our model and dataset construction code will be made available for research purposes.
DATE: September 10, Thursday @ 11:00
Place: EA 516