Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

${C}^{3}$-GS: Learning Context-aware, Cross-dimension, Cross-scale Feature for Generalizable Gaussian Splatting

Created by
  • Haebom

Author

Yuxi Hu, Jun Zhang, Kuangyi Chen, Zhe Zhang, Friedrich Fraundorfer

Outline

This paper aims at generalizable Gaussian splatting, which synthesizes novel views of novel scenes without scene-specific optimization. While existing methods that estimate per-pixel Gaussian parameters using feedforward networks have achieved high-quality synthesis, they struggle to construct accurate geometry from sparse input views. To address this, we propose the $\mathbf{C}^{3}$-GS framework, which enhances feature learning by incorporating context awareness, cross-dimensionality, and cross-scale constraints. $\mathbf{C}^{3}$-GS integrates three lightweight modules to achieve realistic synthesis without additional supervision. Extensive experiments on benchmark datasets demonstrate that $\mathbf{C}^{3}$-GS achieves state-of-the-art rendering quality and generalization ability. The code can be found at https://github.com/YuhsiHu/C3-GS .

Takeaways, Limitations

•
Takeaways:
◦
High-quality image synthesis with accurate geometry even from sparse views.
◦
Realistic compositing possible without additional supervision.
◦
Enhanced feature learning with context-aware, cross-dimensional, and cross-scale constraints.
◦
Achieving state-of-the-art rendering quality and generalization capabilities.
◦
Reproducibility is ensured through open code.
•
Limitations:
◦
Lack of discussion of Limitations of the proposed method.
◦
Further analysis of generalization performance for different types of scenes is needed.
◦
Evaluation of computational cost and memory efficiency is required.
👍
Made with Slashpage