World's Largest Scale: Rvision Open - Sources Indoor 3D Dataset Realsee3D

As the first large model for spatial depth estimation that supports panoramic image input, Argus 1.0 is trained based on the tens of millions of 3D spatial database accumulated by Rushi.

On December 16, 2025, Realsee announced the official release of 10,000 sets of indoor 3D datasets, Realsee3D, for academic research and non-commercial use. This is possibly the world's largest 3D spatial dataset to date, aiming to provide high-quality data for researchers and developers in the field of spatial intelligence, thereby accelerating technological iteration and application implementation across the industry.

Realsee3D

Previously, Realsee released the spatial depth estimation large model Argus 1.0. As the first large model for spatial depth estimation supporting panoramic image input, Argus 1.0 was trained on Realsee's database of tens of millions of 3D spatial data. The Realsee3D dataset released this time consists of high-quality samples carefully selected from this vast database.

Advantages of the Dataset

Realsee3D is a large-scale multi-view RGB-D dataset designed to advance research in indoor 3D perception, reconstruction, and scene understanding. The dataset offers the following advantages:

Large Scale

- 10,000 unique indoor 3D scenes

- 95,962 subdivided room units

- 299,073 sets of viewpoints/RGB-D image pairs

Comprehensive Annotations

To enable multi-task learning, we provide detailed ground truth annotations that extend beyond vision to geometry and semantics:

- Geometry layer: High-precision CAD drawings and floor plans are provided.

- Semantic layer: Includes 2D semantic segmentation and 3D detection labels.

Diverse Scenarios

To ensure the robustness of models in complex real-world environments, we adopted a dual-engine strategy of "real data + procedural generation":

- Real data: 1,000 real-world scenes (capturing complex lighting, layouts, and traces of daily life in the physical world)

- Synthetic data: 9,000 synthetic scenes (based on style templates carefully curated by over 100 professional designers, covering a vast array of furniture models and decoration styles)

Data Types

Color panoramic images
Depth maps
Poses
CAD drawings
Floor plans
Semantic segmentation labels
3D object detection labels

Color panoramic images

Depth maps

Surface normals

Semantic segmentation maps

Applicable Areas

For a long time, research and applications in the field of spatial intelligence have been hampered by a shortage of high-quality spatial data. Leveraging its technical expertise and resource accumulation in 3D spatial data, Realsee is bridging this gap.

This dataset is suitable for core research areas in spatial intelligence such as geometric reconstruction, multi-modal learning, and embodied intelligence. We welcome researchers and developers worldwide to download and use the Realsee3D dataset to explore the future of spatial intelligence research.

How to Access: Currently, the Realsee3D dataset is available for application through official channels. You can access it via Realsee's GitHub repository.