ReLaGS: Relational Language Gaussian Splatting
This repository contains the official Hugging Face model release for ReLaGS (CVPR 2026).
Project page: https://dfki-av.github.io/ReLaGS/
Overview
ReLaGS is a framework for open-vocabulary 3D scene understanding built on top of Gaussian Splatting reconstructions.
It combines:
- A Graph Neural Network (GNN) trained on 3RScan for predicting open-vocabulary relations in 3D scene graphs
- A set of Gaussian Splatting reconstructed scenes enriched with open-vocabulary semantic features
The release is intended for evaluation and reproducibility of open-vocabulary scene graph reasoning in reconstructed 3D scenes.
Repository Structure
1. GNN Model (GNN_model/)
This directory contains a pretrained Graph Neural Network for predicting open-vocabulary relations in 3D scene graphs.
Training data:
- 3RScan dataset
Task:
Given a scene graph with object-level nodes, predict:
- Pairwise relations between objects
- Open-vocabulary relation labels (language-based)
Inputs:
- Node features (geometry + appearance + learned embeddings)
- Scene graph structure (nodes + edges)
Outputs:
- Directed edges with predicted relation labels
2. Scenes (scenes/)
This directory contains reconstructed Gaussian Splatting scenes (based on LeRF-style reconstructions) augmented with semantic features.
Citation
If you find ReLaGS useful for your research, please cite:
@inproceedings{xiearafa2026relags,
title = {ReLaGS: Relational Language Gaussian Splatting},
author = {Xie, Yaxu and Arafa, Abdalla and Javanmardi, Alireza and Millerdurai, Christen and Hu, Jia Cheng and Wang, Shaoxiang and Pagani, Alain and Stricker, Didier},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2026}
}