BROWSE
Detail
Conference Paper
StereoVisPoseNet: Stereo-based Visibility-aware Egocentric 3D Pose Estimation Network
Cited 0 time in 
Share 
Share
- Authors
- Sungjin Hong, Hye-sun Kim, Cho-rong Yu, Youn-hee Gil
- Issue Date
- 2025-11
- Citation
- Symposium on Virtual Reality Software and Technology (VRST) 2025
- Language
- English
- Type
- Conference Paper
- Abstract
- Egocentric 3D pose estimation is challenging due to occlusions and errors at articulated joints. We propose StereoVisPoseNet, a stereo-based visibility-aware network that integrates depth and explicit joint visibility prediction to guide Transformer-based regression and refinement. Our method reduces MPJPE from 76.04 mm to 31.91 mm and PA-MPJPE from 63.43 mm to 28.73 mm compared to UnrealEgo, with substantial improvements for arms and legs. These results demonstrate the importance of combining stereo depth with visibility-aware modeling for robust egocentric pose estimation.
- KSP Keywords
- 3D pose estimation, transformer-based