HKUST-SAIL/Gaussian-Splatting-with-Stochastic-Solids

Geometry-Grounded Gaussian Splatting

What it solves

This project addresses the challenge of grounding the geometry in Gaussian Splatting, enabling the extraction of high-quality meshes from 3D Gaussian representations. It aims to improve the accuracy of reconstructed surfaces compared to standard Gaussian Splatting techniques.

How it works

The system utilizes "Stochastic Solids" to ground the geometry of Gaussian Splatting. It integrates several advanced components to refine the reconstruction process, including:

  • Loss terms from 2DGS and PGSR.
  • A densification strategy from GOF (Gaussian Opacity Fields).
  • Filters from Mip-Splatting.
  • Optional Spherical Gaussian appearance models for improved rendering.
  • A mesh extraction process using Marching Tetrahedra (via CGAL) to convert the learned representation into a usable 3D mesh.

Who it’s for

This tool is designed for researchers and developers in computer vision and graphics who are working on 3D scene reconstruction, novel view synthesis, and high-fidelity mesh extraction from point-based representations.

Highlights

  • Supports high-quality mesh extraction via Marching Tetrahedra.
  • Compatible with standard datasets like DTU and Tanks and Temples.
  • Integrates multiple state-of-the-art techniques for geometry grounding and appearance modeling.
  • Provides a dedicated viewer based on the SIBR framework for visualizing results.

Related

  • Project
    AIGeeksGroup/3D-R1

    3D‑R1 is a research‑grade, open‑source 3‑D vision‑language model that improves reasoning by pre‑training on a synthetic CoT dataset (Scene‑30K) and fine‑tuning with RLHF (GRPO) using perception, semantic‑similarity and format rewards. It dynamically selects informative viewpoints, supports a full training pipeline (SFT + RL), and includes scripts for inference and Rerun‑based visualisation. The repo supplies environment setup, data‑download instructions, and links to pretrained checkpoints and the paper.

  • Project
    prov-gigatime/GigaTIME

    A multimodal AI project that generates virtual spatial proteomics (mIF) profiles from routine H&E pathology slides for tumor microenvironment modeling.

  • Project
    alicevision/CCTag

    A computer vision library for the detection and accurate localization of circular fiducial markers made of concentric circles.

  • Project
    openxrlab/xrmocap

    A PyTorch-based toolbox for multi-view motion capture that provides a unified framework for estimating 3D human poses and shapes from multiple calibrated cameras.