6/n 🛡️How do we ensure no gaps?
Two complementary mechanisms:
1️⃣ Normal alignment loss: aligns rendered normals with image-space depth gradients.
2️⃣ Normal-aware densification: detects gaps in the shell, clones Gaussians with flipped normals, and closes the holes.
5/n 🌊Gaussian Vector & Normal Fields, once Gaussians wrap the scene, the "Objects as Volumes" framework [Miller et al.] yields closed-form expressions for occupancy & normal fields.
The Vector Field V(x) encodes geometry directly from 3DGS params!
7/n Primal Adaptive Meshing (PAM)🕸️
Standard meshing ties vertices to Gaussian positions, so resolution is stuck.
PAM fully decouples mesh resolution from the Gaussians, enabling region-of-interest meshing at arbitrary resolution. 🚀
3/n ⛈️ The root problem: 3DGS treats each Gaussian as a symmetric blob of density.
But surfaces are asymmetric — they separate empty space from occupied matter. Modelling surface points with symmetric primitives introduces a fundamental reconstruction bias.
4/n🔧Our fix: give each Gaussian a learnable oriented normal n. ⬅️
This makes it an oriented stochastic surface element with a clear inside/outside. The Gaussian models density decay on the outward side, while the inward side is treated as fully occupied.
1/n 🧵 Introducing Gaussian Wrapping — a principled framework for extracting high-quality meshes from 3DGS! 🚲
We recover thin structures, like bicycle spokes, where all prior methods fail.
Follow the thread for a brief overview and links!
I’ll be at #SIGGRAPHAsia2025 next week presenting our paper MILo! Join the Neural Fields and Surface Reconstruction session on Tuesday, December 16.
If you’ll be in Hong Kong and would like to discuss research, or grab a coffee ☕️ feel free to reach out.
8/n Huge thanks to my co-authors
@antoine-guedon.bsky.social, Nissim Maruani, @s2.hk, and Maks Ovsjanikov, the supporting institutions, and to the "Objects as Volumes" team for the inspiring framework! 🙏
Check out the full paper and code here:
diego1401.github.io/BlobsToSpoke...
Video
Diego Gomez
Diego Gomez
Diego Gomez
Diego Gomez
Diego Gomez
Diego Gomez
Diego Gomez
Diego Gomez
Diego Gomez
Diego Gomez
1/n🚀Gaussians > Differentiable function > Mesh?
Check out our new work: MILo: Mesh-In-the-Loop Gaussian Splatting!
🎉Accepted to SIGGRAPH Asia 2025 (TOG)
MILo is a novel differentiable framework that extracts meshes directly from Gaussian parameters during training.
🧵👇
