README.md

metadata

license: cc-by-nc-4.0

LRID-simplified Dataset

📖 Overview

This is a simplified version of the Low-light Raw Image Denoising (LRID) dataset, designed for low-light image denoising research. It is part of the PMN_TPAMI project (Learnability Enhancement for Low-light Raw Denoising: A Data Perspective, TPAMI 2024).
Compared to the full LRID dataset, this version retains only the essential components for validation purposes while significantly reducing the size.

GitHub: https://github.com/megvii-research/PMN/tree/TPAMI

🗂️ Dataset Structure

PMN_TPAMI 
 └─LRID_simplified                  # Simplified dataset for training & evaluation
     ├─bias                           # Dark frames (sensor bias)
     ├─bias-hot                       # Dark frames in hot mode of the sensor
     │
     ├─indoor_x5                      # Indoor scenes
     │   │─100                          # Long-Exposure Raw Data (ISO-100)
     │   │   ├─000                      # Scene Number (only the first frame of the original 25 is kept)
     │   │   └─...
     │   │
     │   ├─6400                       # Low-light noisy data (ISO-6400)
     │   │   ├─1                        # Digital gain (low-light ratio)
     │   │   │   ├─000                  # Scene number (only the first frame of the original 10 is kept)
     │   │   │   └─...
     │   │   └─...
     │   │
     │   ├─npy                        # Binary data
     │   │   └─GT_align_ours            # Ground truth after multi-frame fusion
     │   │
     │   ├─ref                        # Long-exposure reference data (ISO-100)
     │   │   ├─000                      # Scene number
     │   │   │   ├── *.dng                # RAW reference frame (ISO-100)
     │   │   │   └── *.jpeg               # JPEG image after camera ISP
     │   │   └─...
     │   │
     │   └─metadata_indoor_x5_gt.pkl  # Metadata (white balance, CCM, etc.)
     │
     ├─outdoor_x3                     # Outdoor scenes
     │   └─...                          # (Structure similar to indoor_x5)
     │
     ├─indoor_x3                      # Indoor scenes with ND filter
     │   └─...                          # (Structure similar to indoor_x5)
     │
     ├─outdoor_x5                     # Outdoor scenes (abandon)
     │   └─...                          # (Structure similar to indoor_x5)
     │
     └─resources                      # Noise calibration results (e.g., dark shading)

🎯 Key Features

• Simplified version:
  • ✅ Only the first noisy frame per scene/setting is kept (originally 10 frames per scene for training).
  • ✅ All ground truth images are retained.
  • ✅ Dark frames (bias, bias-hot) are included for calibration.
  • ✅ Reference data (ref/) provides ISO-100 long-exposure RAW and its camera-processed JPEG for each scene.
  • ❌ Original images used for dataset creation are excluded.
  • ❌ Intermediate results are excluded.
  • ❌ Most noisy frames used for training are excluded.

📊 Scene Categories

Subset	Description	Gain	Lighting Condition
indoor_x5	Indoor scenes	5×	Low-light
outdoor_x3	Outdoor scenes	3×	Low-light
indoor_x3	Indoor scenes with ND filter	3×	Controlled lighting
outdoor_x5 (abandon)	Outdoor scenes	5×	Low-light (a little misalignment)

🔍 Contents Description

• Noisy frames (e.g., indoor_x5/6400/1/000/): contain the first frame of each burst in RAW format.
• Ground truth (npy/GT_align_ours): multi-frame fusion results stored as NumPy binaries (aligned and denoised).
• Reference (ref/): for each scene, an ISO-100 long-exposure RAW and its corresponding JPEG (processed by the camera ISP) are provided as ideal references.
• Metadata (.pkl files): include white balance gains, color correction matrices (CCM), and other camera parameters essential for accurate raw data processing.
• Dark frames (bias/, bias-hot/): sensor bias frames under normal and hot modes, useful for noise calibration and dark current subtraction.
• Resources (resources/): calibration data such as dark shading patterns.

⚙️ Usage Notes

• This dataset is intended for validation or testing rather than full training due to the reduced number of noisy frames.
• Each scene originally contained 10 frames; only the first frame (index 000) is preserved in this simplified version.
• Ground truth images are generated by aligning and fusing multiple frames, providing high-quality clean references.
• The ref/ folders contain the ISO-100 long-exposure captures, which serve as near-ideal clean references for evaluating denoising performance.
• Metadata files are crucial for correctly interpreting the raw sensor data; please refer to them when applying any ISP pipeline.

🔗 Access

Full LRID Dataset: The complete version with all training frames (10 frames per scene) is currently being uploaded and will be available soon.

URL: https://huggingface.co/datasets/hansen97/LRID

📝 Citation

If you use this dataset in your research, please cite:

@inproceedings{feng2022learnability,
    author = {Feng, Hansen and Wang, Lizhi and Wang, Yuzhi and Huang, Hua},
    title = {Learnability Enhancement for Low-Light Raw Denoising: Where Paired Real Data Meets Noise Modeling},
    booktitle = {Proceedings of the 30th ACM International Conference on Multimedia},
    year = {2022},
    pages = {1436–1444},
    numpages = {9},
    location = {Lisboa, Portugal},
    series = {MM '22}
}

@ARTICLE{feng2023learnability,
  author={Feng, Hansen and Wang, Lizhi and Wang, Yuzhi and Fan, Haoqiang and Huang, Hua},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Learnability Enhancement for Low-Light Raw Image Denoising: A Data Perspective}, 
  year={2024},
  volume={46},
  number={1},
  pages={370-387},
  doi={10.1109/TPAMI.2023.3301502}
}

📧 Contact

For questions or issues, please contact Hansen at hansen97@outlook.com.