Abstract
Poker is a high-stakes game involving a deceptive strategy called bluffing and is an ideal research subject for improving high-stakes deception detection (HSDD) techniques like those used by interrogators. Multiple HSDD studies involve staged scenarios in controlled settings with subjects who were told to lie. Scenarios like staged interrogations are inherently poor data sources for HSDD because the subjects will naturally respond differently than someone who actually risks imprisonment, or in the case of poker, loses great sums of money. Thus, unstaged data is a necessity. Unlike traditional HSDD methods involving invasive measurement of biometric data, using video footage of subjects allows for analyzing people’s natural deceptions in real high-stakes scenarios using facial expressions. Deception detection generalizes well for different high-stakes situations, so the accessibility of data in videos of poker tournaments online is convenient for research on this subject. In the hopes of encouraging additional research on real-world HSDD, we present a novel in-the-wild dataset using four different videos from separate professional poker tournaments, totaling 48 minutes. These videos contain great variety in head poses, lighting conditions, and occlusions. We used players’ cards and bets to manually label bluffs and then extracted facial expressions in over 31,000 video frames containing face images from 25 players. We used the dataset to train a state-of-the-art convolutional neural network (CNN) to identify bluffing based on face images, achieving high accuracy for a baseline model. We believe this dataset will allow future in-the-wild bluff detection research to achieve higher deception detection rates, which will enable the development of techniques for more practical applications of HSDD such as in police interrogations and customs inspections.