.png){kind=icon}
.png){kind=icon}
Return
LeanGS-Avatar: region-adaptive densification for efficient reconstruction of head avatars
.png){kind=icon}
PRE
AI
R
W
W
Z
X
DOI：10.1007/s00530-026-02579-1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reconstructing animatable and high-quality 3D human head avatars from monocular video is a core task in applications such as VR/AR and human–computer interaction. Compared to methods based on Neural Radiance Fields (NeRF), 3D Gaussian Splatting (3DGS)-based head avatar reconstruction methods have made significant progress in rendering performance, yet they still suffer from issues such as long training times and poor reconstruction detail when facial expressions vary significantly. To address these challenges, we propose LeanGS-Avatar, a method that optimizes the spatial distribution of Gaussian points by introducing facial region segmentation and dynamic thresholding for Gaussian point densification. Additionally, we introduce a new metric, comprehensive contribution, to prune inefficient Gaussian points, further reducing the number of Gaussians while ensuring high reconstruction quality. To accelerate the reconstruction process, we adopt a first-frame-guided initialization method for Gaussian points, significantly improving model convergence speed. Experimental results show that, compared to existing methods such as SplattingAvatar and GaussianAvatars, which require 554k and 72k Gaussian points respectively, our method only requires 40k Gaussian points to achieve optimal or near-optimal performance in quantitative metrics (PSNR, SSIM, LPIPS), resulting in more realistic rendering quality with a real-time rendering frame rate of up to 210 frames per second.
Keywords:
Monocular video
3D Gaussian splatting
Facial region segmentation
Gaussian densification
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers