A Scalable Sampling Approach for Artificial Intelligence-Based Alcohol Content Estimation in Movies

Area of Science:

• Computational social science
• Media studies
• Public health research

Background:

• Streaming services increase movie accessibility, leading to greater exposure to alcohol portrayals.
• Alcohol depictions in media are a known risk factor for increased alcohol consumption.
• Estimating alcohol exposure in films is challenging due to the extensive time and resources required for frame-by-frame analysis.

Purpose of the Study:

• To evaluate the impact of reduced frame sampling rates on the accuracy of alcohol exposure estimation in movies.
• To determine a practical and computationally efficient sampling frequency for large-scale media analysis.

Main Methods:

• Utilized a LLaVA v1.6 model with 95% accuracy for zero-shot alcohol depiction prediction on 20 feature-length movies.
• Compared full-framerate (25 fps) analysis against uniform downsampling (1 fps) and sparse interval sampling (1 frame per N seconds).
• Quantified accuracy loss using a difference score and measured execution time for each sampling method.

Main Results:

• A sampling frequency of 1 fps achieved an average difference score below 0.10, indicating minimal accuracy loss compared to full-framerate analysis.
• Reducing sampling to 1 fps resulted in a 25-fold decrease in execution time.
• Sparser sampling intervals (e.g., 1 frame per 10 seconds) led to significantly higher error scores.

Conclusions:

• Lowering movie frame sampling frequency to 1 fps offers a practical balance between accuracy and computational efficiency for alcohol exposure studies.
• 1 fps is a scalable solution for estimating alcohol depictions in large movie datasets.
• This method facilitates more efficient public health research on media's influence on alcohol use.