license: cc-by-4.0
pretty_name: Real-world Polarimetric Environment Map Dataset
language:
- en
tags:
- computer-vision
- computer-graphics
- polarization
- stokes
- hdr
- environment-map
- image-based-lighting
size_categories:
- n<1K
configs:
- config_name: full_stokes
data_files:
- split: train
path: full_stokes/metadata.csv
- config_name: linear_stokes
data_files:
- split: train
path: linear_stokes/metadata.csv
Real-world Polarimetric Environment Map Dataset
A real-world dataset of full-sphere HDR RGB polarization environment maps at 2048 × 1024 resolution, with processed EXRs, RAW captures, calibration files and masks.
Project page · Code · Contact
Collections
| Collection | Released scenes | Scene IDs | Measured components |
|---|---|---|---|
full_stokes |
97: 30 indoor, 67 outdoor | 1–100 | S0, S1, S2, S3 |
linear_stokes |
9: 4 indoor, 5 outdoor | 1–9 | S0, S1, S2 |
Linear S3 is not measured; stored zeros are placeholders (s3_measured=false). IDs are local to each collection. Full-Stokes scenes 74, 87 and 98 are withheld for privacy; their metadata rows have empty data paths. Contact the author for access inquiries.
Folder structure
{full_stokes,linear_stokes}/
calibration/
penvmap/{thetaphi,duff,s2l2}/{scene_id}.exr
preview/{s0,s1n,s2n,s3n,aolpn}/{scene_id}.jpg
mask/{saturation,privacy,invalid}/{scene_id}.png
raw/{scene_id}/
metadata.csv
load_scene.py
requirements.txt
README.md
LICENSE
SHA256SUMS
Metadata paths are relative to each collection directory. Use EXRs for numerical analysis; JPEGs are previews. In masks, 255 marks the condition and 0 means unflagged.
Usage
From the dataset root:
pip install -r requirements.txt
python load_scene.py --root . --subset full_stokes --scene-id 1
python load_scene.py --root . --subset linear_stokes --scene-id 1
from load_scene import load_scene
scene = load_scene(".", 1, subset="full_stokes")
stokes = scene["data"] # float32, (1024, 2048, 3, 4): RGB × S0–S3
valid = ~(scene["saturation_mask"] | scene["privacy_mask"] | scene["invalid_mask"])
The default representation is thetaphi; use representation="duff" or "s2l2" to select another. S2L2 arrays have shape (1024, 2048, 3, 12).
License
Released dataset assets, including RAW, processed data and calibration, are licensed under CC BY 4.0; see LICENSE. Withheld data are excluded. Software licensing is separate.
Citation
@inproceedings{oh2026penvmap,
author = {Oh, Yonghee and Maeda, Ryota and Moon, Yunseong and Baek, Seung-Hwan},
title = {Real-world Polarimetric Environment Map Dataset},
booktitle = {SIGGRAPH Asia 2026 Conference Papers},
year = {2026},
doi = {10.1145/3829340.3842262}
}