hbb1/2d-gaussian-splatting

[SIGGRAPH'24] 2D Gaussian Splatting for Geometrically Accurate Radiance Fields

What it solves

This project addresses the lack of geometric accuracy in traditional 3D Gaussian Splatting. By representing scenes as 2D oriented disks (surfels) rather than 3D ellipsoids, it creates radiance fields that are more geometrically precise and easier to convert into high-quality 3D meshes.

How it works

The system uses a set of 2D oriented disks to represent a scene and renders them using a perspective-correct differentiable rasterizer. To improve reconstruction quality, it employs specific regularizations for normal consistency and depth distortion. For mesh extraction, it uses TSDF (Truncated Signed Distance Function) fusion, featuring both a bounded mode for specific volumes and an unbounded mode that contracts space into a sphere with adaptive truncation to handle large-scale scenes.

Who it’s for

This tool is designed for researchers and developers working in computer vision, 3D reconstruction, and novel view synthesis who need high-fidelity geometric surfaces rather than just visually plausible renders.

Highlights

  • Geometrically Accurate: Uses 2D surfels to ensure better surface alignment and accuracy.
  • Differentiable Rasterization: Implements a custom CUDA-accelerated surfel rasterizer for efficient training.
  • Flexible Meshing: Supports both bounded and unbounded mesh extraction for various scene sizes.
  • Broad Compatibility: Works with COLMAP and NeRF Synthetic datasets and integrates with several viewers including SIBR and WebGL-based tools.

Related

  • Project
    zhanghang1989/ResNeSt

    A Split-Attention Network variant of ResNet that improves performance for computer vision tasks like object detection and semantic segmentation.

  • Project
    deepdoctection/deepdoctection

    A Python library for document understanding that orchestrates layout analysis, OCR, and token classification to extract structured data from PDFs and scans.

  • Project
    Peterande/D-FINE

    D‑FINE is a real‑time object‑detection framework that redefines DETR’s box regression as a fine‑grained distribution refinement and adds a self‑distillation module. It offers several model sizes (‑N to ‑X) that achieve state‑of‑the‑art COCO/AP scores (42.8 %–55.8 %) while running at 70‑470 FPS on a T4 GPU, with no extra inference cost. The repo provides pretrained checkpoints, config files, and full training/evaluation scripts for COCO, Objects365, and custom COCO‑style datasets.

  • Project
    opengeos/geoai

    A Python package that integrates AI frameworks with geospatial data analysis, providing tools for satellite imagery processing, model training, and inference.