wildrelight / README.md
Lez's picture
Update README.md
90ab579 verified
metadata
license: cc-by-4.0
task_categories:
  - image-to-image
tags:
  - relighting
  - HDR
  - environment-map
  - outdoor
  - real-world
pretty_name: WildRelight

WildRelight Dataset

A real-world relighting dataset of 30 outdoor/semi-indoor scenes captured across multiple time steps throughout the day. Each scene provides paired HDR photographs and HDR environment maps, enabling evaluation of outdoor relighting tasks.

Download Command

Download small resolution for quick verification.

hf download lez/wildrelight --repo-type dataset --include "small-aligned/**" --local-dir ./wildrelight/small-aligned

Download full resolution dataset.

hf download lez/wildrelight --repo-type dataset --include "ori-aligned/**" --local-dir ./wildrelight/ori-aligned

Key Statistics

Property Value
Scenes 30
Time steps per scene 6–8 (spaced ~1 hour apart)
Total HDR photo/envmap pairs 196
Photo resolution (full) 6048 × 4024 px
Envmap resolution (full) 7680 × 3840 px (equirectangular)
Photo resolution (small) 1200 × ~798 px
Envmap resolution (small) 1200 × 600 px
File format OpenEXR
HDR normalization Radiance at reference shutter speed (1/500 – 1/10 s, per scene)

Equipment

  • Main camera: Sony α7 series — RAW
  • 360° camera: Insta360 Pro2 — 6-lens equidistant fisheye, stitched to equirectangular

Dataset Variants

Two resolution variants are provided under ori/ and small/. Within each variant, every scene contains four folders:

{variant}/{scene}/
├── photo/              # HDR photograph (unaligned)
├── photo_aligned/      # HDR photograph (temporally aligned, recommended)
├── envmap/             # HDR environment map (unaligned)
└── envmap_aligned/     # HDR environment map (temporally aligned, recommended)

Both photo_aligned/ and envmap_aligned/ are temporally aligned to the first time step — each image is warped so that the static background is registered across time. Use these for relighting experiments.

File Naming

photo/          time{N}_hdr.exr          (N = 0, 1, 2, …)
envmap/         time{N}_envmap.exr

Metadata

Each scene contains a meta.json with:

Field Description
scene Scene name
ref_shutter Reference shutter speed (seconds) used for HDR normalization
ref_shutter_fraction Human-readable, e.g. "1/500"
hdr_method HDR merge algorithm (triangular_weight_debevec1997)
alignment_method life_optical_flow or cv2_sift_homography
gps {latitude, longitude, altitude_m} — available for 10 scenes
photos[].shooting_time Capture timestamp from EXIF
photos[].iso ISO speed
photos[].shutter_speed Representative shutter speed (human-readable)
envmaps[].shooting_time Capture timestamp
envmaps[].shutter_speed Envmap shutter speed

Scene List

GPS coordinates are provided as additional information. Since collecting GPS coordinates was not part of the original plan, only 10 scenes contain GPS coordinates.

Scene Time steps Alignment GPS
bench2 6 LIFE flow
bench3 6 LIFE flow
bike 6 LIFE flow yes
bikes2 6 LIFE flow
bridge 6 LIFE flow
building304 6 LIFE flow
building324 7 LIFE flow
byway 6 LIFE flow yes
cabin 6 CV2/SIFT
corridor 6 LIFE flow yes
foodtruck 6 LIFE flow yes
garden 6 LIFE flow
grassland 7 LIFE flow yes
grassland2 6 LIFE flow yes
indoor 6 CV2/SIFT
lake 6 LIFE flow
meadow 6 LIFE flow
parking 7 LIFE flow
pillar 6 LIFE flow
playground 7 LIFE flow
rail 7 LIFE flow
road 7 LIFE flow
sand 6 CV2/SIFT
seaside 8 CV2/SIFT
stairs 6 LIFE flow yes
statue 6 LIFE flow yes
statue2 6 LIFE flow yes
table_tennis 6 LIFE flow yes
wall 6 LIFE flow
yatai 7 LIFE flow

Temporal Alignment

Photos are aligned to time0 of each scene. Two methods are used depending on scene content:

  • LIFE optical flow (26 scenes): dense flow-based warping, handles gradual lighting changes and small camera motion
  • CV2/SIFT homography (4 scenes: cabin, indoor, sand, seaside): feature-based homography, used when flow alignment fails due to detailed textures.

Alignment is computed independently for ori and small resolutions. Environment maps are aligned using phase-correlation-based rotation estimation.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

If you use this dataset, please cite:

@misc{wang2026wildrelight,
      title={WildRelight: A Real-World Benchmark and Physics-Guided Adaptation for Single-Image Relighting}, 
      author={Lezhong Wang and Mehmet Onurcan Kaya and Siavash Bigdeli and Jeppe Revall Frisvad},
      year={2026},
      eprint={2605.11696},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.11696}, 
}