ttgroup/blueneg-release

BlueNeg: A 35mm Negative Film Dataset for Restoring Channel-Heterogeneous Deterioration

Description

This set of images is collected for studying the research problem of restoring the corrupted negative films. Due to the physical nature of negative film, the red, green, and blue light-sensitive layers are located differently inside the negative film. Therefore, the rates of deterioration of these three layers are different. Tien-Tsin Wong found that the blue channel is relatively vulnerable compared to the other two channels. Because the blue light-sensitive layer is on the outermost layer on the emulsion side. That is, the deterioration rates are heterogeneous. This characteristic is inherited from its ancestor, i.e., storing the three color negatives (R, G, B) on three separate glass plates by Prokudin-Gorsky.

Since the blue channel is more vulnerable and the other two channels are relatively well-preserved, this means we can restore the blue channel by exploiting the retained information from the red and green channels to restore the color photograph. This is especially sound with the latest AI technologies. Unfortunately, most existing photo restoration techniques are developed based on printed photographs, in which the nature of deterioration is different from that of the negatives. This is why this dataset is created.

Data Information

Data Source	Partition	Number	Folder	Bits per channel	Resolution (may varies)	File extension	Public
Negative films	blue-corrupted	297	negative-16bit	16	10,128 x 6,840	dng	Yes
	blue-intact	194
Printed photo	blue-corrupted	247	printed-16bit	16	2,994 x 1,920	tif	Yes
	blue-intact	151
Negative preview (after negation)	blue-corrupted	297	negative-preview-8bit	8	1,322 x 892	preview.png	Yes
	blue-intact	194
Pseudo ground truth	blue-corrupted	247	pseudogt-8bit	8	< 1,322 x 892	pseudogt.png	Yes
Printed photo (testset)	blue-corrupted	30	-	16	2,994 x 1,920	tif	No
Pseudo ground truth (testset)	blue-corrupted	30	-	8	< 1,322 x 892	pseudogt.png	No

Meta Information

The meta information is stored in the meta.json and transformations.pkl. The meta information includes the following fields:

• meta.json
  • filename
  • partition
  • is_testset
  • date
  • roll_id
  • number_in_roll
  • location
  • geo_location
  • film_type
  • negative_path
  • printed_path
  • preview_path
  • pseudogt_path
  • scene_property
    • is_indoor
    • is_daytime
• transformations.pkl: a python dict using filename (without extension) as key, and the value is a dict with the following fields:
  • matrix: the perspective projection matrix to warp the printed photo to the negative preview
  • bbox: the bounding box of the warped printed photo in the negative preview; the bounding box is in the format of [x0, y0, x1, y1], preview[y0:y1, x0:x1, :] has the same size as pseudo ground truth

Naming Convention

YYYYMMDD[R]-NN-NAME.EXT

• YYYYMMDD: The date when the photo was taken (might not always be accurate, but roughly correct).
• [R]: the roll number, if available
• NN: the photo number in the roll
• NAME: the photo's name, usually the location or the subject of the photo.
• EXT: the file extension, e.g., .dng, .tif, .preview.png, .pseudogt.png, etc.

License

Our license shares the same spirit with the CC-BY 4.0. However, please note that our license is not CC-BY 4.0. Please check the LICENSE file for more details.

All photos are owned and copyrighted by Tien-Tsin Wong. You are 
automatically granted with permission to use the images for academic 
and commerical usages, provided that the image credit 
"Copyrighted by Tien-Tsin Wong" is included in any forms of 
publication, reproduction, redistribution, or derivatives of the images.

Citation

If you find the dataset is useful to you, please also cite our publication below in your work/publication.

• Hanyuan Liu, Chengze Li, Minshan Xie, Zhenni Wang, Jiawen Liang, Chi-Sing Leung, and Tien-Tsin Wong, "BlueNeg: BlueNeg: A 35mm Negative Film Dataset for Restoring Channel-Heterogeneous Deterioration", arXiv preprint, 2024.

@misc{blueneg,
      title={BlueNeg: A 35mm Negative Film Dataset for Restoring Channel-Heterogeneous Deterioration}, 
      author={Hanyuan Liu and Chengze Li and Minshan Xie and Zhenni Wang and Jiawen Liang and Chi-Sing Leung and and Tien-Tsin Wong},
      year={2024},
      eprint={to be updated},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}