Towards Face De-identification for Wearable Cameras

Wearable cameras provide valuable new sources of data for health and wellness monitoring, however, such visual data brings privacy concerns. This paper proposes a prototype egocentric face de-identification system for wearable camera images by swapping the original faces with synthetic faces. The motivation of this paper is to: (1) de-identify faces in egocentric images and (2) preserve the existence of each identity in images where the source identity is altered. The system incorporates our proposed method, which promises a privacy-aware and cost-effective approach. We evaluated the system on the Ego4D audio-visual PoV diarization training set by analysing six activities where faces are visible in wearable camera data. The results show promising de-identification on the source faces while most existences remain.