--- pretty_name: SkyLume license: other viewer: false gated: true tags: - 3d - image - geospatial - computer-vision - aerial-imagery - uav - 3d-reconstruction - novel-view-synthesis - inverse-rendering - lidar - slam task_categories: - image-to-3d extra_gated_heading: "Request access to SkyLume" extra_gated_description: "SkyLume is released for approved academic research use. Please provide enough information for the maintainers to review your request." extra_gated_button_content: "Agree and request access" extra_gated_prompt: "By requesting access, you agree to use SkyLume only under the stated research data use terms, not redistribute the dataset or share access credentials, and cite the associated paper and dataset repository." extra_gated_fields: Full name: text Affiliation / Institution: text Institutional email: text Country: country Intended use: type: select options: - Academic research - Reproducibility study - Education - Other Project title: text Brief research purpose: text I agree to use the dataset for non-commercial research only: checkbox I agree not to redistribute the dataset or share my access token: checkbox I agree to cite the associated paper and this dataset repository: checkbox --- # SkyLume ![SkyLume teaser](assets/teaser1.jpg) SkyLume is a large-scale real-world UAV dataset for urban scene reconstruction under varying illumination. It captures the same urban regions across morning, noon, and afternoon flights, pairing high-resolution five-direction UAV imagery with LiDAR-derived geometry for robust 3D reconstruction, novel view synthesis, inverse rendering, and cross-time consistency research. Project page: [https://zhuoxiaoli.github.io/skylume_page/](https://zhuoxiaoli.github.io/skylume_page/) Paper: [Beyond a Single Light: A Large-Scale Aerial Dataset for Urban Scene Reconstruction Under Varying Illumination](https://arxiv.org/abs/2512.14200) ## Dataset Highlights - 10 urban regions covering approximately 7.21 square kilometers. - 109K+ UAV images captured across 3 illumination periods. - 6252 x 4168 per-view image resolution. - Five-direction aerial imagery: one nadir view and four oblique views. - LiDAR-supported geometry, including meshes and depth supervision. - Repeated RTK-guided flights for cross-time reconstruction and consistency evaluation. The complete Hugging Face release is expected to be about 3 TB. The upload is being completed in batches; check the **Files** tab for the current set of available archives. ## Sensors and Capture Protocol SkyLume was collected with a DJI M350 RTK UAV, a CHCNAV C30 five-direction aerial camera, and a DJI L2 LiDAR unit. Each region was flown repeatedly along comparable trajectories at three daily illumination periods, with approximately 80 percent forward overlap, 60 percent side overlap, around 120 m flight height, and 1 Hz camera triggering. ## Released Modalities The release is designed to support both appearance and geometry evaluation. Depending on the scene package, it includes: - COLMAP-style SfM packages. - Per-period camera poses. - Mesh-derived depth. - LiDAR-derived depth. - Mesh normals. - Post-processed meshes. - Solar-angle annotations. ## File Organization The dataset is distributed as large ZIP archives. The upload is incremental, so not every archive may be available yet. ### Single-Period Packages | Path | Approx. archive size | | --- | ---: | | `single_period/buildings.zip` | 7.65 GB | | `single_period/campus_full_penta_cam.zip` | 53.4 GB | | `single_period/campus_full_single_cam_period_1.zip` | 70.2 GB | | `single_period/campus_full_single_cam_period_2.zip` | 33.0 GB | | `single_period/campus_full_single_cam_period_3.zip` | 42.5 GB | | `single_period/main_campus.zip` | 30.6 GB | | `single_period/moon_bay.zip` | 63.6 GB | | `single_period/town.zip` | 21.4 GB | ### Multi-Period Packages | Path | Approx. archive size | | --- | ---: | | `multi_period/gym.zip` | 97.1 GB | | `multi_period/ipark.zip` | 59.7 GB | | `multi_period/main_campus.zip` | 129 GB | | `multi_period/staff_residence.zip` | 83.3 GB | | `multi_period/tec_school.zip` | 37.2 GB | Checksums are stored in `checksums/sha256sums.txt` as each archive is registered and uploaded. ## Download SkyLume is a gated dataset. Request access from this dataset page and wait for approval. After approval, log in with the Hugging Face CLI: ```bash hf auth login ``` Download selected files instead of cloning the entire repository: ```bash hf download vavaw/SkyLume \ --type dataset \ --include "multi_period/tec_school.zip" \ --include "checksums/sha256sums.txt" \ --local-dir ./SkyLume ``` For a different archive, replace the `--include` pattern with the path shown in the file table or in the **Files** tab. ## Checksum Verification After downloading, verify the ZIP archive against the published SHA256 checksum. On Linux: ```bash cd SkyLume grep "multi_period/tec_school.zip" checksums/sha256sums.txt | sha256sum -c - ``` On macOS: ```bash cd SkyLume grep "multi_period/tec_school.zip" checksums/sha256sums.txt | shasum -a 256 -c - ``` The checksum confirms that the downloaded file matches the version registered for this Hugging Face release. ## Access and Use Access is intended for non-commercial academic research, reproducibility studies, and education. Users should not redistribute the archives, mirror the dataset, share access tokens, or use the data outside the approved research purpose. See [DATA_USE_AGREEMENT.md](DATA_USE_AGREEMENT.md) for the data use terms associated with this release. ## Citation If you use SkyLume, cite the associated paper and this dataset repository: ```bibtex @misc{li2026singlelightlargescaleaerial, title={Beyond a Single Light: A Large-Scale Aerial Dataset for Urban Scene Reconstruction Under Varying Illumination}, author={Zhuoxiao Li and Wenzong Ma and Taoyu Wu and Jinjing Zhu and Shuai Zhang and Jing OU and Tongyan Hua and Yinrui Ren and Rongjun Qin and Hui Xiong and Wufan Zhao}, year={2026}, eprint={2512.14200}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2512.14200} } ``` ## Maintenance Notes - The upload is in progress and the file table should be treated as the planned release structure until all archives are present. - The public preview link on the project page is intentionally not repeated here; this Hub repository is the controlled distribution entry point. - If a file is missing from the **Files** tab, it has not yet completed the batch upload and verification process.