zju3dv/LoG

Level of Gaussians

What it solves

LoG addresses the challenge of creating highly realistic, real-time renderable models of urban-scale environments (covering millions of square meters) using accessible hardware, such as a single RTX 4090 GPU.

How it works

The project leverages PyTorch and Gaussian Splatting techniques to synthesize views of large-scale scenes. It uses Colmap for dataset preparation and automatically configures heuristic parameters based on the size of the dataset to optimize training.

Who it’s for

Researchers and developers working on 3D scene reconstruction, urban-scale view synthesis, and real-time rendering of large environments.

Highlights

  • Urban-Scale Capability: Designed specifically for scenes spanning millions of square meters.
  • Hardware Efficiency: Capable of training and rendering on a single consumer-grade RTX 4090 GPU.
  • Real-Time Rendering: Focuses on high-fidelity, real-time view synthesis.
  • Automated Configuration: Automatically adjusts parameters based on dataset size for easier training.

Related

  • Project
    cs-chan/Total-Text-Dataset

    A word-level English curved text dataset and evaluation framework for improving the detection of arbitrary-shaped text in natural scene images.

  • Project
    zju3dv/DetectorFreeSfM

    A detector-free Structure from Motion (SfM) system that reconstructs 3D scenes from images without relying on traditional feature detectors, achieving top performance in the Image Matching Challenge 2023.

  • Project
    weecology/DeepForest

    A Python package for training and predicting ecological objects, such as tree crowns and birds, in airborne RGB imagery using deep learning object detection.

  • Project
    huawei-bayerlab/marigold-v2

    Marigold V2 is a research codebase that repurposes a pretrained diffusion transformer (Qwen‑Image‑Edit‑2509) into fast, single‑step predictors for monocular depth, see‑through depth, surface normals, and albedo. It offers cheap fine‑tuning (≈5 days on a single 32 GB GPU), state‑of‑the‑art benchmark results, ready‑to‑use inference/evaluation scripts, and a modular YAML‑driven training framework that can be extended to new dense‑vision tasks.