MrKarimiDandClaude Opus 4.8 395ea78bec Route per-image masks into per-camera folders in openSFM processing
ns-process-data lantern-openSFM expects each camera's masks at
<camera>/mask/<image_stem>.png, but the PanDORA dataset ships them in a flat
data/masks/ folder. Add _populate_camera_mask_folder() to copy each mask into
the per-camera mask/ dir before projection (handles masks/<camera>/<stem>.png,
masks/<camera>_<stem>.png, and masks/<stem>.png source layouts), and make the
mask output dir creation robust with mkdir(parents=True, exist_ok=True).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-08-02 15:31:54 -04:00
2024-05-19 16:34:05 -04:00
2023-06-05 09:56:07 -07:00
2023-08-12 20:17:38 +00:00
2023-08-15 17:34:23 -04:00
2023-05-18 07:13:58 -07:00
Exp
2024-05-18 19:48:08 -04:00
…
2025-07-10 19:27:31 -04:00

PanDORA: Casual HDR Radiance Acquisition of Indoor Scenes for Image-based Lighting

Official code release for the paper "PanDORA: Casual HDR Radiance Acquisition of Indoor Scenes for Image-based Lighting" (ICCP 2026).

PanDORA reconstructs an HDR radiance field of an indoor scene from two 360° videos captured at different exposures. This repo lets you download a scene and train the model with a single Docker environment.

Built on Nerfstudio. The hdr-nerfacto method is our implementation of HDR-NeRF (Huang et al., CVPR 2022).


Quick start

Requirements: an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit.

Everything runs inside a Docker image that this repo builds for you.

1. Build the image and open a shell

docker/build_and_run.sh --shell

The first build takes a while (it compiles a few components). After it finishes you land in a shell inside the container. All the commands below run there.

2. Download a scene

docker/download_scenes.sh --list                     # see all available scenes
docker/download_scenes.sh --dest /data meeting_room  # download one (~5 GB)

3. Prepare the scene

ns-process-data lantern-openSFM \
  --data /data/meeting_room/data/ \
  --output-dir /data/meeting_room/meeting_room_ns/ \
  --metadata /data/meeting_room/sfm/reconstruction.json

4. Train

DATASET_DIR=/data/meeting_room/meeting_room_ns/ docker/run_pandora_pipeline.sh

Results are written under outputs/pandora/PanDORA/.

Tip: for a quick test run, add STEP1_ITERS=200 STEP2_ITERS=200 before the command to train only a few iterations.


Advanced

The steps above cover the common case. For the full workflow — capturing your own scenes, running OpenSFM, and computing metrics — see:


Citation

@misc{dastjerdi2025pandoracasualhdrradiance,
  title         = {PanDORA: Casual HDR Radiance Acquisition of Indoor Scenes for Image-based Lighting},
  author        = {Mohammad Reza Karimi Dastjerdi and Dominique Tanguay-Gaudreau and Frédéric Fortier-Chouinard and Yannick Hold-Geoffroy and Nima Kalantari and Jean-François Lalonde},
  year          = {2025},
  eprint        = {2407.06150},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2407.06150}
}

This work builds on Nerfstudio and, for hdr-nerfacto, HDR-NeRF — please cite them too:

@inproceedings{nerfstudio,
  title        = {Nerfstudio: A Modular Framework for Neural Radiance Field Development},
  author       = {Tancik, Matthew and Weber, Ethan and Ng, Evonne and Li, Ruilong and Yi, Brent and Kerr, Justin and Wang, Terrance and Kristoffersen, Alexander and Austin, Jake and Salahi, Kamyar and Ahuja, Abhik and McAllister, David and Kanazawa, Angjoo},
  year         = 2023,
  booktitle    = {ACM SIGGRAPH 2023 Conference Proceedings},
  series       = {SIGGRAPH '23}
}

@inproceedings{huang2022hdr,
  title        = {HDR-NeRF: High Dynamic Range Neural Radiance Fields},
  author       = {Huang, Xin and Zhang, Qi and Feng, Ying and Li, Hongdong and Wang, Xuan and Wang, Qing},
  booktitle    = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages        = {18398--18408},
  year         = {2022}
}
Languages
Python 91.2%
JavaScript 6.8%
TypeScript 0.8%
Shell 0.4%
SCSS 0.4%
Other 0.4%