Spaces:

---

CompVis

/

EDGS

like 32

Running on Zero

TLDR:

EDGS changes 3D Gaussian Splatting by eliminating the traditional densification process, enabling 30-second scene reconstruction from real-world videos. Our demo showcases this breakthrough capability, reconstructing complex scenes at unprecedented speeds while maintaining quality.

• EDGS replaces iterative densification with direct correspondence-based initialization
• Achieves 10× faster convergence than standard 3DGS
• Maintains compatibility with existing 3DGS pipelines

Abstract

3D Gaussian Splatting reconstructs scenes by starting from a sparse Structure-from-Motion initialization and iteratively refining under-reconstructed regions. This process is inherently slow, as it requires multiple densification steps where Gaussians are repeatedly split and adjusted, following a lengthy and redundant optimization path. Moreover, this incremental approach often leads to suboptimal renderings, particularly in complex or high-frequency regions where detail is critical.
We propose a fundamentally different approach: eliminating the densification step and replacing it with a direct initialization of 3D Gaussians using triangulated pixels from dense image correspondences. This initialization provides each Gaussian with well-informed colors, scales, and positions from the outset, dramatically shortening the optimization path and removing the need for densification. Unlike traditional methods that rely on sparse keypoints, our dense initialization ensures uniform detail across the scene, even in high-frequency regions where 3DGS and other methods typically struggle. Additionally, all splats are optimized in parallel from the start, bypassing the need to wait for densification to add new Gaussians.
EDGS not only outperforms speed-optimized models in training efficiency but also achieves higher rendering quality than state-of-the-art approaches, all while using only half the splats of standard 3DGS. It is fully compatible with other 3DGS acceleration techniques, making it a versatile and efficient solution that can be integrated with existing approaches.

Our demo allows us to reconstruct scenes on real video in around 30 seconds. One of the examples:

This grant will help us:
✅ Democratize access to real-time 3D reconstruction
✅ Provide pre-configured Spaces for immediate testing
✅ Support educational content for the 3D vision community