A Real-world Display Inverse Rendering Dataset

Seokjun Choi1, Hoon-Gyu Chung1, Yujin Jeon1 Giljoo Nam2, Seung-Hwan Baek1
1POSTECH 2Meta Reality Lab
Paper arXiv Code Data

TL;DR

We present a novel real-world display inverse rendering dataset, DIR. The DIR dataset consists of:

  • 144 images of 16 objects captured under OLAT lighting
  • Ground truth surface normal map
  • Polarization RAW signals
  • Stereo pair images
  • Carefully-calibrated light informations

We provide the dataset in an organized format, along with benchmark results for photometric stereo, single-view & multi-light inverse rendering, and baseline code for display inverse rendering. Please refer to our GitHub page for details.

Gallery

Figure 1

Overview

The capture system consists of an LCD monitor as the light source and a stereo polarized camera for imaging. For each object, we sequentially projected 144 (16 × 9) OLAT patterns and captured them. The scanned mesh was then aligned to the scene pose to obtain ground truth depth and normal maps.

Reference Image (0)

Results (Photometric Stereo)

Avg. Elephant Owl Cat Frog Robot Pig Chicken Girl Boy Nefertiti Trex Gnome Horse Snowman Plaster Objet
Woodham 23.14 27.02 26.60 21.05 21.58 28.18 17.02 18.39 24.86 21.44 37.03 18.98 19.83 19.27 32.21 19.56 17.28
PS-FCN 17.29 20.26 15.17 10.61 19.15 16.68 15.80 11.91 25.96 22.27 20.03 18.22 19.33 17.48 18.75 17.25 7.73
SRSH 25.25 26.21 18.49 16.95 23.42 19.09 32.76 17.88 37.14 31.19 23.97 25.05 27.44 27.70 27.96 26.93 21.87
SCPS-NIR 24.17 22.75 7.93 8.97 16.28 17.87 34.89 10.43 45.12 37.18 52.97 21.85 16.64 48.98 15.65 21.30 7.94
UniPS (M=64) 23.27 24.93 18.33 19.54 24.99 22.18 25.72 23.07 26.38 30.65 28.71 21.86 24.48 26.72 18.89 19.43 16.39
SDM-UniPS (M=64) 14.89 18.83 14.37 9.70 14.12 14.85 15.33 16.05 14.99 15.22 22.73 14.58 13.46 16.93 15.18 12.55 9.38

Qualitative Results
Figure 3
Figure 4
Figure 5

Citation

@article{choi2025real,
  title={A Real-world Display Inverse Rendering Dataset},
  author={Choi, Seokjun and Chung, Hoon-Gyu and Jeon, Yujin and Nam, Giljoo and Baek, Seung-Hawn},
  journal={ICCV},
  year={2025}
}