Topology-Aware Optimization of Gaussian Primitives for Human-Centric Volumetric Videos

SIGGRAPH Asia 2025


Yuheng Jiang1,2,3*, Chengcheng Guo2*, Yize Wu2, Yu Hong2, Shengkun Zhu2, Zhehao Shen2, Yingliang Zhang4, Shaohui Jiao3, Zhuo Su3, Lan Xu2†, Marc Habermann1†, Christian Theobalt1†

1Max Planck Institute for Informatics, Saarland Informatics Campus
2ShanghaiTech University
3ByteDance Inc.
4DGene Digital Technology Co., Ltd.

Paper Video Code Dataset Unity Plugin

Given a dynamic sequence with complex topological changes, TaoGS can not only achieve high-fidelity reconstruction and real-time rendering, but also supports up to 40x compression.

Overview Video

We present TaoGS, a novel topology-aware dynamic Gaussian representation that disentangles motion and appearance to support, both, long-range tracking and topological adaptation with up to 40x compression.

Pipeline

We propose a novel motion-to-appearance Gaussian representation for robust tracking and high-fidelity rendering of general 4D scenes with topological changes. We track sparse motion Gaussians and incorporate new candidate Gaussians through a spatial-temporal tracker and error map to model new observations. The motion Gaussians are then transformed into a Gaussian Look-Up Table (GLUT), activating corresponding appearance Gaussians, which can be packed into 2D attribute maps for efficient video codec compression.

Dataset

Overview of our dataset. The first two rows show representative static samples from our multi-view capture system. The bottom row displays consecutive frames from a dynamic sequence, illustrating the topological changes we address.

Comparison

Qualitative comparison with SOTA methods on novel view synthesis on HiFi4G dataset.
Ours
GT
Ours
4DGS
Ours
Spacetime Gaussian
Ours
DualGS
Ours
HiFi4G
Ours
V3

Result Gallery

Acknowledgements

The authors would like to thank Meihan Zheng and Yiwen Cai from ShanghaiTech University for processing the dataset. We also thank the reviewers for their feedback. This work was supported by National Key R&D Program of China (2022YFF0902301), Shanghai Local college capacity building program (22010502800). We also acknowledge support from Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI).