Synapse
⌘+K
Synapse
PulseExploreClubsResearchersJournals
Instagram
HomeClubsExplore
March 26, 2026Complex & Intelligent SystemsOpen Access

End-to-end high-fidelity avatar reconstruction using 3D Gaussian splatting with monocular video as the sole input

View Full Paper
Ask AI
Bookmark
Share

Authors

TFTeng FeiQXQin XinJDJ. G. Deng

Discussion

Loading...

Member takes

Overview

End-to-end approach shows improved avatar reconstruction quality in augmented and virtual reality, suggesting a simpler method for realistic representation.

Key Points

  • This research aims to improve 3D avatar reconstruction using monocular video input by addressing challenges in accuracy and realism.
  • Introduced a tile-based rasterization method within the 3D Gaussian Splatting framework.
  • Utilized Gaussian ellipsoids initialized through Patch Match Stereo techniques.
  • Employed active and passive optimization for refining Gaussian parameters and density filtering.
  • Achieved high-fidelity avatar reconstructions without visual artifacts.
  • Outperformed traditional avatar reconstruction techniques and existing radiation field methods.
  • Demonstrated enhanced flexibility and a lightweight pipeline for avatar generation.

Cite This Study

Fei et al. (2026) studied this question.

synapsesocial.com/papers/69c4cd05fdc3bde448918d12https://doi.org/10.1007/s40747-026-02231-5
View Full Paper
Ask AI
Bookmark
Share