Core Concepts
MediaPipe Holistic (MPH) is not reliable for eyebrow movement analysis in sign languages, requiring additional correction models.
Abstract
1. Introduction
Deep Learning advancements enhance Computer Vision (CV) solutions.
OpenPose and OpenFace are popular CV packages used in various applications.
MediaPipe Holistic (MPH) tracks body, hands, and facial landmarks in 2D video data.
2. Methods
Analyzed two data sets: Kazakh-Russian Sign Language (KRSL) utterances and new head tilts/eyebrow raises data set.
MPH tested against OpenFace (OF) for eyebrow position tracking.
3. Results
MPH results show similarities with corrected OF outputs but have issues with inner and outer eyebrow distances.
Head movements distort MPH outputs significantly.
4. Discussion
MPH's distortion during head movements makes it unsuitable for linguistic analysis without correction models.
Corrective models may not perform as well due to the complexity of MPH distortions.
5. Bibliographical References
Lists references related to sign language analysis using Computer Vision tools.
Stats
MPHの出力は、頭部運動によって著しく歪む。
MPHの結果は、修正されたOFの出力と類似しているが、内側と外側の眉間距離に問題がある。