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>
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).
- 📄 Paper: https://arxiv.org/abs/2407.06150
- 🌐 Project page: https://lvsn.github.io/pandora/
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=200before 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:
README-pandora.md— the PanDORA pipeline in detail.README-hdrnerfacto.md— the HDR-Nerfacto pipeline.
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}
}