Feature Splatting: Language-Driven Physics-Based Scene Synthesis and Editing

1UC San Diego, 2MIT 3IAIFI

*Indicates equal contribution

ECCV 2024

TLDR: Splatting features enables automatic physics-based scene editing

Abstract

Scene representations using 3D Gaussian primitives have produced excellent results in modeling the appearance of static and dynamic 3D scenes. Many graphics applications, however, demand the ability to manipulate both the appearance and the physical properties of objects. We introduce Feature Splatting, an approach that unifies physics-based dynamic scene synthesis with rich semantics from vision language foundation models that are grounded by natural language. Our first contribution is a way to distill high-quality, object-centric vision-language features into 3D Gaussians, that enables semi-automatic scene decomposition using text queries. Our second contribution is a way to synthesize physics-based dynamics from an otherwise static scene using a particle-based simulator, in which material properties are assigned automatically via text queries. We ablate key techniques used in this pipeline, to illustrate the challenge and opportunities in using feature-carrying 3D Gaussians as a unified format for appearance, geometry, material properties and semantics grounded on natural language.

New: Try Feature splatting in NerfStudio✨!

Fast feature extraction speed (~0.2s per image)🐇.

Feature splatting trains right away - taking only ~5mins for installation, weight download, and feature extraction!

In-browser interactive scene editing.

We implemented custom widgets in NerfStudio to support in-browser geometric editing and physics simulation based on feature splatting.

Language-Driven Physics-Based Scene Synthesis


Feature splatting jointly optimizes for a unified Gaussian representation for appearance, geometry, and semantics. Feature splatting goes beyond open-vocabulary segmentation to dive into component-level details to automatically determines physical properties of materials based on features.


Shake a rigid vase with elastic flower stems.

Make a slim standing statue move like jelly.

Turn a lego bulldozer into a pile of sand.

Make a ball fall off from a chair.

Simulate physics effects on gaussian-ized synthetic assets from Objaverse.

Geometric Editing


Feature splatting implements several editing primitives for scene editing, such as object removal, scaling, rotation, translation, and cloning.

Method Overview


Splatting Features onto Gaussians


We fuse features from multiple foundation 2D Vision models. We train latent features to rasterize both SAM-enhanced CLIP features and DINOv2 features for better quality.

PCA Visualization of latent features of every Gaussian in the garden vase scene using Vuer.


Optimized Feature Rasterization


Feature splatting is engineeringly optimized to support future Gaussian+feature research. We implemented custom CUDA kernels for optimized memory access pattern to rasterize high-dimensional features.


MPM-based Physics Simulation


Given an object segmented by user query, Feature splatting uses a pre-defined set of names of rigid material queries (e.g., ceramic and wood) to determine physical property of components, and estimate ground a support using language queries (e.g., floor and table). These information is turned in to a customized physics engine based on Taichi for physics simulation.


Overview Video

BibTeX

@inproceedings{qiu-2024-featuresplatting,
      title={Language-Driven Physics-Based Scene Synthesis and Editing via Feature Splatting},
      author={Ri-Zhao Qiu and Ge Yang and Weijia Zeng and Xiaolong Wang},
      booktitle={European Conference on Computer Vision (ECCV)},
      year={2024}
    }