Yihalem Yimolal Tiruneh

Yihalem Yimolal Tiruneh

M.S. Researcher · Vision & Learning Lab, UNIST. 3D human avatar reconstruction, 3D Gaussian splatting, and vision–language models.

Unified 3D Gaussian Human Avatars Under Arbitrary Body Visibility

Published in European Conference on Computer Vision (ECCV) 2026, 2026

Authors. Y. Y. Tiruneh, M. S. Ali, U. Jeong, M. A. Khan, MD K. C. Sayem, A. Bayramgeldiyev, B. Bhattarai, S. Baek.

Venue. Accepted to ECCV 2026.

Project page. yihalem1.github.io/FlexiAvatar

Summary.

  • Proposed a visibility-aware optimization framework that restricts Gaussian splatting supervision to observed body regions, eliminating hallucinated geometry and texture drift under partial-view inputs.
  • Unified pipeline supporting full-body, upper-body, and head-only avatars with occlusion-robust SMPL-X tracking, FLAME-based facial initialization, part-specific residual MLPs, and diffusion-based texture completion.
  • Achieved state-of-the-art on NeuMan, ZJU-MoCap, TalkShow, and INSTA; reduced memory by ~50% and improved FPS by ~34% for head-only avatars; ~3% PSNR gain overall.