Estimating Intermediate Alcohol Intoxication States from Paired Facial Images

Seamless, effortless assessment of personal alcohol responses is important for supporting safe, enjoyable drinking. Although breath alcohol concentration (BrAC) is a key indicator, it is unsuitable for continuous monitoring because it requires active user participation. Similarly, subjective ratings require repeated user input and are difficult to obtain continuously in natural drinking situations. Recent vision-based methods have achieved progress; however, they primarily focus on binary intoxication detection and do not address the gradual progression of subjective intoxication. We address this limitation by developing a deviation-based estimation framework that evaluates a person’s intoxication state by comparing their appearance after alcohol consumption with their sober appearance as the reference state. By modeling how an individual deviates from their usual state, the system produces scores that reflect intermediate intoxication levels. Although the model was trained solely with binary labels and without subjective-rating supervision, its scores were associated with subjective intoxication ratings. This finding indicates that graded, subjective-aligned structure can emerge from readily attainable binary supervision. These results suggest the practical value of deviation-based modeling in capturing intermediate alcohol intoxication states.