Reproduce “4D Scaffold Gaussian Splatting with Dynamic-Aware Anchor Growing for Efficient and High-Fidelity Dynamic Scene Reconstruction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Woong Oh Cho, In Cho, Seoha Kim, Jeongmin Bae, Youngjung Uh, Seon Joo Kim
A 4D anchor-based Gaussian Splatting framework reconstructs dynamic scenes with high fidelity while significantly reducing storage requirements. The method achieves a PSNR of 28.86 on dynamic regions of the N3DV dataset, outperforming prior 4DGS methods while reducing storage by over 97%.
Key insights
- Retaining a sufficient number of Gaussians while compressing their parameters via an anchor-based system is more effective for high-fidelity dynamic scene reconstruction than simply reducing the Gaussian count.
- Effective resource allocation in dynamic scenes requires anchor growing strategies that specifically account for the transient nature of dynamic elements, such as weighting gradient signals by temporal coverage.
- Specialized and compact parameterizations for neural 4D Gaussians, including linear motion and a generalized Gaussian for temporal opacity, significantly enhance their expressiveness and efficiency in modeling spatiotemporal changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.