InteriorGS: 3D Gaussian Splatting Dataset of Semantically Labeled Indoor Scenes

A comprehensive indoor scene dataset featuring 3D Gaussian representations with semantic annotations and spatial occupancy information.

Sample from the InteriorGS dataset. The dataset provides high-quality 3D Gaussian Splatting (3DGS) representations along with instance-level semantic bounding boxes and occupancy maps indicating agent-accessible areas. The red and yellow trajectories denote the paths of a ground robot and a UAV (drone), respectively. Notably, InteriorGS supports free-form agent navigation and interaction within a continuous 3D environment, enabling realistic spatial intelligence training and evaluation.

🤖RGB rendering and corresponding depth map from the perspective of the human/ground robot along the red trajectory

🚁RGB rendering and corresponding depth map from the UAV (drone) perspective along the yellow trajectory.

🔄 News

• 2025-07-24: Initial release InteriorGS (v1.0) with 500 scenes.

📋 Overview

Existing 3D indoor scene datasets mainly fall into two categories, each with its own limitations. Datasets reconstructed from RGB-D scans are typically represented as triangle meshes, but often suffer from incomplete geometry, occlusions, and missing semantic or structural annotations. Due to the incompleteness of the reconstructed scenes, they do not support continuous agent movement, limiting their use in navigation and interaction tasks.

In contrast, artist-designed 3D scenes provide high visual quality, but are available in limited numbers and often lack consistent semantic and spatial annotations, making them less suitable for large-scale scene understanding.

To overcome these issues, we present InteriorGS, a dataset of 1,000 indoor scenes represented using 3D Gaussian Splatting (3DGS). We create it by photorealistically rendering handcrafted 3D environments, then reconstructing splatting-based representations from over 5 million images, resulting in high-quality and lightweight 3D data that supports real-time rendering. Each scene is semantically annotated at the object level, with more than 554,000 object instances across 755 categories, and each object is paired with a 3D bounding box. We also provide an occupancy map for each scene to support navigation and spatial understanding.

InteriorGS includes over 80 types of indoor environments, such as homes, convenience stores, wedding banquet halls, and museums. It supports various downstream tasks, including 3D scene understanding, controllable scene generation, and embodied agent navigation, and offers a valuable resource for research in spatial intelligence.

We provide a sample viewer for viewing our data.

🗂️ Dataset Structure

InteriorGS
├── 0001_839920             # 3D Gaussian scene with detailed information
│   ├── 3dgs_compressed.ply # Gaussian point cloud files
│   ├── labels.json         # Semantic annotations and bounding boxes
│   ├── occupancy.png       # Grayscale occupancy map
│   └── occupancy.json      # Occupancy metadata
├── 0002_839955
│   ├── 3dgs_compressed.ply
│   ├── labels.json
│   ├── occupancy.png
│   └── occupancy.json
└── ...

📊 Data Description

3D Gaussian Models (.ply files)

• Format: PLY (Polygon File Format)
• Content: 3D Gaussian parameters including position, covariance, opacity, and spherical harmonics coefficients. The data is compressed using the SuperSplat method to optimize storage and loading efficiency.
• Coordinate System: The coordinate system of the scenes is defined as XYZ = (Right, Back, Up)
• Units: Meters

Semantic Annotations

• Bounding Boxes: 3D oriented bounding boxes defined by 8 corner vertices
• Labels: Per-object semantic class labels with instance IDs
• Format: JSON files containing arrays of annotated objects

Occupancy Maps

• Resolution: 1024x1024 pixels (configurable)
• Format: PNG (grayscale) + JSON (metadata). Metadata is used for mapping between the 2D pixels and 3DGS world coordinates.
• Coverage: Top-down view of navigable floor space
• Values:
  • 255 (white): Free space
  • 0 (black): Occupied space
  • 127 (gray): Unknown space

🏠 Citation

If you use InteriorGS in your research or development, please cite or link to our project page:

@misc{InteriorGS2025,
  title        = {InteriorGS: A 3D Gaussian Splatting Dataset of Semantically Labeled Indoor Scenes},
  author       = {SpatialVerse Research Team, Manycore Tech Inc.},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/spatialverse/InteriorGS}}
}

📄 License

This dataset is released under InteriorGS License.

🙏 Acknowledgements

This project makes partial use of code from the SuperSplat repository. We thank the authors for open-sourcing their implementation.