SeokjunChoi/snapshot-polarimetric-dir

Snapshot Polarimetric Display Inverse Rendering

Pretrained weights and sample captures for Snapshot Polarimetric Display Inverse Rendering, ACM Transactions on Graphics 45(6), Proc. SIGGRAPH Asia 2026.

Code: https://github.com/MichaelCSJ/PDIR

Files

File	Size	Contents
pdir_best.ckpt	329 MB	joint synthetic + real checkpoint, step 14250
sample_data.tar.gz	26 MB	five real captures: cat, bowl, case, foil, owl

Each sample directory holds exactly what the network reads:

<scene>/
  mask.png
  quad{0,1,2,3}_main_hdr_{s0,s1,s2}.npy    per-quadrant Stokes, [384, 384, 3] float32

Usage

git clone https://github.com/MichaelCSJ/PDIR.git && cd PDIR
pip install -r requirements.txt
python download_assets.py

python inference.py --ckpt checkpoints/pdir_best.ckpt \
    --dataset real --data-root sample_data --out-dir results \
    --exposure-norm p95_0.95

--exposure-norm p95_0.95 is required: it is the per-scene HDR normalisation this checkpoint was trained with, and inference has to match it. The checkpoint also expects --pattern-color-strength 1.0, which is the default.

Citation

@article{choi2026snapshot,
  title     = {Snapshot Polarimetric Display Inverse Rendering},
  author    = {Choi, Seokjun and Moon, Yunseong and Kang, Kaizhang and Chung, Hoon-Gyu
               and Kim, Jin-Nyeong and Nam, Giljoo and Baek, Seung-Hwan},
  journal   = {ACM Transactions on Graphics},
  volume    = {45},
  number    = {6},
  articleno = {201},
  year      = {2026},
  doi       = {10.1145/3842531}
}