Article

Evaluation of Anonymized Sign Language Videos Filtered Using MediaPipe

This study investigates the feasibility of using MediaPipe to anonymize sign language videos. Recent research has developed techniques for anonymizing the identity of a signer in a video, while preserving the signed message. Many of these prototypes are computationally intensive and are not currently useable for everyday automated real-time use. This gap MediaPipe, a tool developed by Google for tracking body movement in video, could be feasible for real-time anonymization, but has not yet been evaluated for its feasibility in sign anonymization. We fill this gap with a study in which deaf signers (n=10) view two filters developed using MediaPipe: a face mesh filter that covers only the face with an avatar-like face mask and a silhouette filter that covers the whole body in a solid monochrome, with interconnected dots showing the skeleton of the signer. Results show that signers are adept at understanding and reproducing short sentences covered by either filter. However, the filters are described as unnatural, and signers note facial movements are limited. We conclude that MediaPipe is likely robust enough for understanding manual information in signs but not necessarily for capturing facial information, and we suggest further improvements to the two filters.

Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.