ageppert/single_photon_challenge_full_preprocessed_adaptive

Single Photon Challenge — Full Preprocessed Dataset

Preprocessed measurement/target PNG pairs derived from the Single Photon Challenge reconstruction dataset.

Source

The raw dataset (~425GB training, ~42GB test) is hosted by the WISION Lab at UW-Madison. Photoncubes contain 1024 binary frames from a simulated single-photon camera, paired with ground-truth RGB reconstructions.

• Challenge website: https://singlephotonchallenge.com/
• Download page: https://singlephotonchallenge.com/download
• VisionSIM toolkit: https://visionsim.readthedocs.io/

Preprocessing pipeline

Each photoncube was preprocessed using adaptive similarity-flow-sum registration:

1. Unpack the last 256 binary frames from each photoncube
2. Partition frames into non-overlapping registration blocks of size 8
3. Register each block to the reference (last block) using global scale+translation search over candidates [0.9, 0.94, 0.98, 1.0, 1.02, 1.06, 1.1] with phase cross-correlation (overlap threshold = 0.45)
4. Refine alignment with dense TVL1 optical flow (use_dense_flow=True, attachment=15, tightness=0.3, num_warp=5)
5. Warp and accumulate all frames per accepted block with per-pixel validity masking
6. Invert SPC response → linear RGB flux via flux = -log(1 - p) / 0.5
7. sRGB tonemap → standard gamma curve
8. Save as uint8 PNG

Measurements and targets are stored as 800×800 RGB PNGs.

Dataset statistics

Split	Measurements	Targets	Paired
train	1850	1850	yes
test	185	0	no (test set has no ground truth)
total	2035	1850

Directory structure

single_photon_challenge_full_preprocessed_adaptive/
  metadata.json
  train/
    <scene>/<frame>_measurement.png
    <scene>/<frame>_target.png
  test/
    <scene>/<frame>_measurement.png

Usage

from huggingface_hub import snapshot_download

# Download the full preprocessed dataset
root = snapshot_download(
    repo_id="ageppert/single_photon_challenge_full_preprocessed_adaptive",
    repo_type="dataset",
)

# Or use with the diffusion training codebase:
# Set in config.py:
#   PREPROCESSED_DATA_CONFIG["dataset_source"] = "hf"
#   PREPROCESSED_DATA_CONFIG["dataset_hf_repo"] = "ageppert/single_photon_challenge_full_preprocessed_adaptive"

Preprocessing parameters

{
  "source": "Single Photon Challenge reconstruction dataset",
  "source_url": "https://singlephotonchallenge.com/download",
  "algorithm": "adaptive_similarity_flow_sum",
  "K": 256,
  "reg_block_size": 8,
  "scale_candidates": [
    0.9,
    0.94,
    0.98,
    1.0,
    1.02,
    1.06,
    1.1
  ],
  "overlap_threshold": 0.45,
  "max_global_mse": null,
  "use_dense_flow": true,
  "flow_attachment": 15,
  "flow_tightness": 0.3,
  "num_warp": 5,
  "invert_response": true,
  "invert_factor": 0.5,
  "tonemap": true,
  "split": "all",
  "notes": "Measurements are preprocessed from raw photoncubes using: adaptive block-wise scale+translation registration with optional dense optical-flow refinement, followed by SPC response inversion and sRGB tonemapping. Saved as uint8 PNGs. Targets are copied from original ground-truth PNGs."
}

Citation

If you use this dataset, please cite the Single Photon Challenge:

@misc{singlephotonchallenge,
    title={The Single Photon Challenge},
    author={Jungerman, Sacha and Ingle, Atul and Nousias, Sotiris and Wei, Mian and White, Mel and Gupta, Mohit},
    year={2025},
    url={https://singlephotonchallenge.com/}
}