SuperNormal Data
Data for SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration (CVPR 2024) by Xu Cao and Takafumi Taketomi, CyberAgent.
Paper (arXiv) Β· Code (GitHub)
Files
| File | Size | Contents | Used by |
|---|---|---|---|
data.zip |
1.8 GB | Preprocessed inputs (multi-view normal maps, masks, NeuS-format cameras) for 5 DiLiGenT-MV objects and 3 of our captured objects | download_data.sh, run_diligent.sh, run_own_object.sh |
flower_girl.zip |
5.5 GB | Raw iPhone captures of the flower girl object | data capture & preprocessing guide |
lion.zip |
5.5 GB | Raw iPhone captures of the lion object | same as above |
dog.zip |
5.4 GB | Raw iPhone captures of the dog object | same as above |
SHA256SUMS |
β | SHA-256 checksums of the zip files | β |
Only data.zip is needed to reproduce the results in the paper. The raw captures are for running the full preprocessing pipeline yourself.
data.zip
data/
βββ diligent_mv_normals/ # bear, buddha, cow, pot2, reading (20 views each)
β βββ <object>/
β βββ cameras_sphere.npz
β βββ mesh_Gt.ply
β βββ mask/
β βββ normal_camera_space_sdmunips/
β βββ normal_world_space_GT/
β βββ normal_world_space_sdmunips/
β βββ normal_world_space_sdps/
β βββ normal_world_space_omnidata/
βββ own_objects_normals/ # dog1, lion, woman (18 views each)
βββ <object>/
βββ cameras_sphere.npz
βββ normal_map_all.png
βββ mask/
βββ normal_camera_space_sdmunips/
βββ normal_world_space_sdmunips/
For the DiLiGenT-MV objects, the normal_world_space_* folders hold normal maps from different sources (ground truth or estimated by different methods). Pick one with dataset.normal_dir in the config.
Raw captures (flower_girl.zip, lion.zip, dog.zip)
<object>/
βββ RAW/ # 234 DNG images (18 views Γ 13 images)
βββ mask/ # 18 foreground masks, one per view
βββ cameras.xml
RAW/: DNG images captured with the iPhone's built-in camera app. Each view has one ambient-light image followed by 12 images lit by an active light source at different positions.mask/: foreground masks for each view.cameras.xml: camera parameters calibrated with Agisoft Metashape.
See the preprocessing guide for turning these into the inputs SuperNormal uses.
Download
With the Hugging Face CLI:
pip install -U huggingface_hub
# Preprocessed data (~1.8 GB)
hf download cyberagent/SuperNormal data.zip --repo-type dataset --local-dir .
unzip data.zip
# Raw captures (optional, ~5.5 GB each)
hf download cyberagent/SuperNormal flower_girl.zip --repo-type dataset --local-dir .
hf download cyberagent/SuperNormal lion.zip --repo-type dataset --local-dir .
hf download cyberagent/SuperNormal dog.zip --repo-type dataset --local-dir .
Or from Python:
from huggingface_hub import hf_hub_download
path = hf_hub_download("cyberagent/SuperNormal", "data.zip", repo_type="dataset")
To check a download, put SHA256SUMS next to the zips and run sha256sum -c --ignore-missing SHA256SUMS.
License
The data we captured and processed is released under the MIT License, the same as the code.
The DiLiGenT-MV objects in data/diligent_mv_normals/ (including the ground-truth meshes and normal maps) are derived from the DiLiGenT-MV benchmark by Li et al. Use of that part is subject to the original dataset's terms. Please cite their paper as well if you use it.
Citation
@inproceedings{supernormal2024cao,
title={SuperNormal: {N}eural Surface Reconstruction via Multi-View Normal Integration},
author={Cao Xu and Taketomi Takafumi},
booktitle={CVPR},
year={2024}
}
If you use the DiLiGenT-MV objects:
@article{li2020multi,
title={Multi-View Photometric Stereo: A Robust Solution and Benchmark Dataset for Spatially Varying Isotropic Materials},
author={Li, Min and Zhou, Zhenglong and Wu, Zhe and Shi, Boxin and Diao, Changyu and Tan, Ping},
journal={IEEE Transactions on Image Processing},
volume={29},
pages={4159--4173},
year={2020}
}
- Downloads last month
- 40