Michael-Jiang/See-the-Invisible-with-SWIR

See the Invisible with SWIR Dataset

[Paper] [Code]

Author-owned SWIR noisy/clean pairs and calibration measurements accompanying the ICCV 2025 workshop paper. The canonical public archive contains 20 scenes and 600 pairs from Camera 1, plus five scenes and 350 pairs from Camera 2.

Download the complete release with the official checksum-verifying helper:

pip install "huggingface-hub>=0.25"
python scripts/download_artifacts.py --kind dataset

Contents

Camera	Sensor	Resolution	Scenes	Noisy/clean pairs
Camera 1	Sony IMX990	1280×1024	20	600
Camera 2	Sony IMX991	640×512	5	350

The 36.34 GB release contains 16,678 indexed files:

cameraN/
  benchmark/{noisy_corrected,noisy_original,clean_corrected,clean_original}/
  calibration/{dark_frames,fpn-or-derived,base_k.npy}/
manifests/{camera1_test.csv,camera2_test.csv}
files.csv
files.sha256
dataset_info.json

Camera 1 calibration includes 6,000 dark frames at the six gains used for training. Camera 2 includes 8,000 dark frames, flat-field captures, and derived calibration arrays. SID training images are not redistributed.

File representation

Benchmark images are lossless floating-point TIFFs in calibrated sensor units; negative values after fixed-pattern subtraction are valid and must not be clipped on load. Camera 1 images are 1024×1280 and Camera 2 images are 512×640. Each camera provides noisy_corrected, noisy_original, clean_corrected, and clean_original directories. Calibration folders contain the raw dark/flat captures and derived arrays required by the release code.

Load TIFFs without integer conversion or clipping:

import imageio.v3 as iio

image = iio.imread("camera1/benchmark/noisy_corrected/example.tiff")
assert image.dtype.name == "float64"

Use files.sha256 to verify the download and the CSV manifests for canonical corrected pairs. Original variants reproduce the no-FPN ablation.

Limitations, license, and citation

The packaged manifests define the canonical evaluation selection. Results are camera-specific and may not generalize to other sensors, temperatures, or spectral bands.

Author-owned files are licensed under Creative Commons Attribution 4.0 International. Attribute Haiyang Jiang, Hongjun Wang, and Yinqiang Zheng and cite the accompanying paper. Full terms are linked from LICENSE. SID retains its original third-party terms and is not included.

@inproceedings{jiang2025see,
  title={See the Invisible with SWIR: Learning to Enhance via Synthetic Noise Modeling},
  author={Jiang, Haiyang and Wang, Hongjun and Zheng, Yinqiang},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops},
  pages={7449--7458},
  year={2025}
}