Krause, A. F., Ferger, A., Pitsch, K. (2023). Anonymization of Persons in Videos of Authentic Social Interaction: Model Selection and Parameter Optimization. In: 10th International Conference on CMC and Social Media Corpora for the Humanities (CMC-Corpora 2023), Mannheim.

Title

Krause, A. F., Ferger, A., Pitsch, K. (2023). Anonymization of Persons in Videos of Authentic Social Interaction: Model Selection and Parameter Optimization. In: 10th International Conference on CMC and Social Media Corpora for the Humanities (CMC-Corpora 2023), Mannheim.

Abstract

Automatic anonymization of persons in video recordings requires robust detection of face and head areas. Machine learning-based face and posture detectors provide bounding boxes of face and head regions, but specific parameters need to be optimized to maximize the number of correctly anonymized persons and minimize manual annotation and verification efforts. Three different, state-of-the-art ML models (RetinaFace Detector (RFD), Dual-Shot Face Detector (DSFD) and Yolo7-Pose Detector (Y7PD)) were evaluated regarding their suitability for face- and head-region anonymization. Results on our specific anonymization test dataset show that RFD slightly outperforms DSFD if recall (maximizing anonymization) is favored over precision (minimizing false positive face detections). Y7PD yields an even better recall, but at the cost of comparatively low precision. Besides anonymization, collected detector outputs can provide useful data for multimodal interaction research, like body-posture trajectories and face locations.