Salsa Zufar Radinka Akmal, Naufal Afkaar, Ananda Ravi Kuntadi, Irawati Nurmala Sari, Tirana Noor Fatyanosa
Indonesian Sign Language (BISINDO) serves as the primary visual communication system for Indonesia's deaf community, yet technological support remains limited compared to well-studied sign languages like American Sign Language (ASL). Accurate pose estimation is important for BISINDO applications including sign language recognition and production, but no systematic framework comparison exists to guide researchers in selecting appropriate tools for BISINDO-specific challenges such as twohanded signs, hand-face interactions, and dense facial expressions. This study addresses this gap by conducting the first quantitative comparison of MediaPipe Holistic and Real-Time Multi-Person Pose Estimation (RTMPose) on a curated 1,020-clip BISINDO dataset recorded with native signers. MediaPipe employs a multistage BlazePose architecture generating 543 landmarks with zero-coordinate suppression for low-confidence detections, while RTMPose uses a top-down CSPNeXt approach producing 133 keypoints with continuous output regardless of confidence. We evaluated both frameworks using spatial accuracy metrics-Mean Per Joint Position Error (MPJPE) and Percentage of Correct Keypoints (PCK)-as well as temporal consistency analysis and expert validation from native BISINDO signers. Confidence-aware filtering at threshold 0.5 reduced localization error by 43.5% while retaining 92.7% of keypoints. RTMPose demonstrated superior temporal stability, while MediaPipe received higher ratings from deaf evaluators due to better semantic preservation. This study provides evidence-based guidance for BISINDO pose estimation framework selection and contributes a systematically generated skeleton dataset for BISINDO research. © 2026 IEEE.
Universitas Brawijaya, Department of Informatics Engineering, Malang, Indonesia
