PanoWorld RealSee3D Evaluation Data

This dataset repository contains the RealSee3D evaluation data package used for PanoWorld experiments. The data is distributed as a single versioned archive:

• RealSee3D_eval_data.tar.gz

Dataset Summary

The package contains indoor synthetic scene evaluation data organized by scene and viewpoint. Each scene directory contains scene-level evaluation metadata and a set of viewpoints. Each viewpoint contains a panorama image, six perspective RGB images, corresponding depth maps, radial depth maps, and camera metadata.

Basic Statistics

• Number of scenes: 50
• Number of viewpoints: 615
• Number of files after extraction: 14,245
• Uncompressed source directory size: approximately 6.2 GB
• Compressed archive size: 5,601,252,161 bytes
• SHA256:

42a967717b7740f3c22bfb418da4ee2bdf9db110524c115f3e65ff9cd312aec3

File type counts after extraction:

Directory Structure

After extracting the archive, the data is organized as:

RealSee3D_eval_data/
  synthetic_scene_00415/
    map_eval.json
    map_eval_12.json
    viewpoints/
      0/
        panoImage_1600.jpg
        depth_image.png
        depth_scale.txt
        extrinsics.txt
        transforms.json
        images/
          0.png
          1.png
          2.png
          3.png
          4.png
          5.png
        depths/
          0_depth.png
          1_depth.png
          ...
          5_depth.png
        depths_radial/
          0_depth.png
          1_depth.png
          ...
          5_depth.png
  synthetic_scene_00788/
    ...

The archive root contains the top-level RealSee3D_eval_data/ directory.

File Description

• map_eval.json: scene-level evaluation metadata for the 8-view evaluation setting. The viewpoints included in each key and its corresponding value belong to the same room, while viewpoints associated with different keys and their corresponding values belong to different rooms.
• map_eval_12.json: scene-level evaluation metadata for the 12-view evaluation setting. The viewpoints included in each key and its corresponding value belong to the same room, while viewpoints associated with different keys and their corresponding values belong to different rooms.
• viewpoints/<view_id>/panoImage_1600.jpg: equirectangular panorama image, typically 1600 x 800.
• viewpoints/<view_id>/images/*.png: perspective RGB images, typically 800 x 800.
• viewpoints/<view_id>/depth_image.png: equirectangular depth image, typically 1600 x 800, 16-bit grayscale.
• viewpoints/<view_id>/depths/*_depth.png: perspective depth maps, typically 800 x 800, 16-bit grayscale.
• viewpoints/<view_id>/depths_radial/*_depth.png: perspective radial depth maps.
• viewpoints/<view_id>/depth_scale.txt: scalar depth scale. In the checked sample, the value is 2500.0.
• viewpoints/<view_id>/extrinsics.txt: camera extrinsic matrix for the viewpoint.
• viewpoints/<view_id>/transforms.json: camera intrinsics, image size, depth scale, and per-frame transform metadata.

For example, a checked sample transforms.json contains camera parameters such as fl_x = 400.0, fl_y = 400.0, cx = 400.0, cy = 400.0, w = 800, h = 800, and per-frame paths to images/, depths/, and depths_radial/.

Download

Using the Hugging Face CLI:

huggingface-cli download JiaJinrang/PanoWorld RealSee3D_eval_data.tar.gz \
  --repo-type dataset \
  --local-dir .

Using Python:

from huggingface_hub import hf_hub_download

archive_path = hf_hub_download(
    repo_id="JiaJinrang/PanoWorld",
    filename="RealSee3D_eval_data.tar.gz",
    repo_type="dataset",
)
print(archive_path)

Extraction

tar -xzf RealSee3D_eval_data.tar.gz

Optionally verify the archive before extraction:

shasum -a 256 RealSee3D_eval_data.tar.gz

The expected SHA256 is:

42a967717b7740f3c22bfb418da4ee2bdf9db110524c115f3e65ff9cd312aec3

Intended Use

This package is intended for evaluating PanoWorld-style panoramic indoor 3D reconstruction and novel-view synthesis pipelines. The archive format is kept intact to preserve the original directory layout and make experimental reproduction straightforward.

Limitations

The dataset is provided as a compressed archive rather than decomposed Hugging Face datasets records. As a result, the Hugging Face Dataset Viewer may not preview individual samples directly. Users should download and extract the archive before using the data.

License and Terms

No explicit license file is included in this dataset package at the time of publication. Please use the data according to the terms provided by the dataset owner and any relevant RealSee3D/PanoWorld project agreements. If an official license is available, this section should be updated accordingly.

Citation

If you use this evaluation data, please cite the associated PanoWorld paper:

@misc{jia2026panoworldgenerativespatialworld,
      title={PanoWorld: A Generative Spatial World Model for Consistent Whole-House Panorama Synthesis},
      author={Jinrang Jia and Zhenjia Li and Yijiang Hu and Yifeng Shi},
      year={2026},
      eprint={2605.17916},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.17916},
}

Please also cite the RealSee3D dataset:

@misc{Li2025realsee3d_data,
  doi = {10.5281/zenodo.17826243},
  url = {https://doi.org/10.5281/zenodo.17826243},
  author = {Li, Linyuan and Wu, Yan and Li, Xi and Wang, Lingli and Rao, Tong and Zhou, Jie and Pan, Cihui and Hui, Xinchen},
  title = {Realsee3D: A Large-Scale Multi-View RGB-D Dataset of Indoor Scenes (Version 1.0)},
  publisher = {Zenodo},
  year = {2025}
}

Version

• Initial Hugging Face dataset upload: 2026-05-25
• Archive file: RealSee3D_eval_data.tar.gz