[CVPR2024] HiFi4G: High-Fidelity Human Performance Rendering via Compact Gaussian Splatting

in this work we propose high-i 4G a totally explicit and compact gaussian based approach for High Fidelity 4D Human Performance rendering from dense footage our explicit representation enables the seamless integration of dynamic Human Performance into static scenes we showcase side by side the stable gaussian ellipsoids and jitter-free rendering results our pipeline effectively recovers spatial temporally consistent 4D [Music] gaussians here are our more results our approach delivers real-time High Fidelity rendering of human performance at high resolution High 54g effectively handles complex motions such as playing instruments dancing and changing [Music] clothes [Music] I furthermore we demonstrate the human performance rendering on VR headsets pico4 our compact 4D gaussian representation enables immersive viewing as well as real-time interaction and editing of human performances hii 4G achieves High photo realism and low storage [Music] demands [Music] we also showcase a long sequence of nearly 2,000 frames and multiple performers in the gaussian viewer on a laptop we first compare Hi-Fi 4G with instant NSR and noose 2 Instant NSR suffers from severe artifacts while noose 2 produces blurry [Music] results we then compare our method with human RF and dynamic 3D gaussians human RF exhibits over smoothing on high frequency details Dynamic 3D gaussians fail to recover detailed appearance and view dependency in contrast Hi-Fi 4G surpasses these existing methods delivering detailed and highquality Human Performance rendering we conduct an oblation study on the Dual graph and regularization our full pipeline effectively balances non- rigid prior and gaussian updating ensuring High Fidelity rendering without jitteriness post compression thanks for watching

Channel: ShanghaiTech Digital Human

Views: 8,923

Keywords:

Id: 917WVr2EHh4

Length: 4min 9sec (249 seconds)

Published: Wed Dec 06 2023