cvg/resplat

[ECCV'26 Spotlight] ReSplat: Learning Recurrent Gaussian Splatting

What it solves

ReSplat addresses the challenge of efficiently predicting 3D Gaussian splats from images. It aims to reduce the number of Gaussians required for high-quality reconstruction while providing a mechanism to improve the scene representation during test-time through iterative refinement.

How it works

ReSplat uses a feed-forward recurrent model. It starts with a compact initialization that predicts Gaussians in a subsampled space, using significantly fewer Gaussians than previous per-pixel methods. It then employs a weight-sharing recurrent module that takes the rendering error as a gradient-free feedback signal to iteratively predict updates for the Gaussian parameters.

Who it’s for

This project is for researchers and developers working in 3D scene reconstruction, computer vision, and neural rendering, specifically those utilizing Gaussian Splatting techniques.

Highlights

  • Compact Initialization: Predicts Gaussians in a subsampled space, using 16× fewer Gaussians than prior per-pixel methods.
  • Recurrent Refinement: Uses a weight-sharing module to refine Gaussians based on rendering error without needing gradients.
  • Test-time Adaptation: Allows for iterative improvement of the 3D representation during the inference process.

Related

  • Project
    inspatio/querysplat

    QuerySplat is a 3D Gaussian Splatting prediction framework that decouples geometry and appearance representations to improve 3D scene reconstruction from images.

  • Project
    Linketic/CityGaussian

    A framework for high-quality, large-scale 3D scene reconstruction using Gaussian Splatting, featuring multi-GPU support and geometrically accurate rendering.

  • Project
    zju3dv/InfiniSplat

    InfiniSplat is a framework for large-baseline monocular view synthesis that reconstructs 3D Gaussian Splatting scenes from a single RGB image or an RGB-depth pair.

  • Project
    nerfstudio-project/gsplat

    A CUDA-accelerated library for the rasterization of 3D Gaussians, providing a faster and more memory-efficient way to render radiance fields compared to the official implementation.

  • Dispatch
    Introduction to 3D Gaussian Splatting

    3D Gaussian Splatting is a rasterization technique that enables real-time rendering of photorealistic 3D scenes learned from a small set of images.