MDRS: Multi-Degradation Remote Sensing benchmark, clean training tiles

This upload contains the clean (ground truth) training tiles used to train restoration models for the MDRS benchmark.

What is in here

File	Size	Contents
mdrs_train_clean.tar	1.99 GiB	5,130 PNG tiles under train/clean/
manifest_sha256.txt	375 KiB	Per-file SHA256, one line per tile
CHECKSUMS.txt	192 B	SHA256 and byte count of the tar

Tile properties, verified over the full set:

• Count: 5,130
• Format: PNG, RGB, 8 bit
• Size: 512 x 512, uniform across all tiles
• Naming: 1.png through 5130.png, contiguous, no gaps
• Uncompressed total: 2,123,923,150 bytes

Important: file names collide across splits

The training split and the benchmark test split both number their tiles from 1. So train/clean/1.png and the test split's 1.png are different images. We checked this directly: the two sets share 270 file names and zero identical file contents by MD5.

Do not merge the two splits into one flat directory. Keep the train/ prefix.

How to use

# download
huggingface-cli download --repo-type dataset zzqsb/MDRS \
  mdrs_train_clean.tar manifest_sha256.txt CHECKSUMS.txt --local-dir .

# verify the archive before extracting
sha256sum -c <(grep -v '^#' CHECKSUMS.txt)

# extract, gives ./train/clean/*.png
tar -xf mdrs_train_clean.tar

# verify every tile
cd train/clean && sha256sum -c ../../manifest_sha256.txt

The archive is uncompressed on purpose. PNG is already a compressed format, so gzip would save under 2 percent while costing time on both ends.

What these tiles are for

MDRS evaluates restoration under stacked degradations. The benchmark grid applies degradations cumulatively across six depths, from a single degradation up to six at once, over 37 depth and combination cells.

The six degradation types are blur, cloud, dark, haze, noise, and super-resolution downscaling.

These clean tiles are the training ground truth. Degraded training pairs are generated from them by applying the degradation chain. The benchmark test split is a separate set of tiles and is not included in this upload.

Two properties to know before you use these

Both were measured on the tiles in this upload, not carried over from another split.

1. Low dynamic range, which makes PSNR unusually sensitive

Per-tile pixel standard deviation, in [0,1] units, over a 395 tile sample:

	value
median	0.0618
p10	0.0427
p90	0.0879
natural-image benchmarks such as Set5 or Urban100, for reference	about 0.22

These tiles occupy roughly 28 percent of the dynamic range of a typical natural-image benchmark. PSNR uses a fixed data range of 1.0 and does not normalise by image variance, so a global brightness or contrast shift costs far more here than it would on natural images:

global mean shift	natural image	these tiles
0.05	0.2 sigma	0.8 sigma
0.10	0.5 sigma	1.6 sigma
0.20	0.9 sigma	3.2 sigma

If you are porting a method that was tuned on natural images, expect any step that adjusts global radiometry to be much more damaging here than you are used to.

2. Tile ids are not a random permutation

Neighbouring ids tend to come from the same source scene. Mean feature distance between tiles at a given id gap, relative to the distance between two randomly chosen tiles, measured over ids 1 to 600:

id gap	relative distance
1	0.631
2	0.650
5	0.657
10	0.697
20	0.756
50	0.971

The correlation only washes out around gap 50. Do not take a contiguous id range as a subset. A contiguous block of 1,000 tiles covers far fewer distinct scenes than 1,000 tiles sampled with a stride. If you need a subset, use a stride or a fixed permutation.

Provenance

These are 512 x 512 crops prepared as the clean training corpus for the MDRS benchmark, which targets remote sensing imagery. The benchmark test split is a separate, disjoint set of tiles and is not part of this upload.

The exact upstream source of the crops is not restated here because we have not re-verified it end to end from this working copy. If you need the original source for a citation, please open a discussion on this repo and we will confirm it rather than guess.

Citation

If you use this data, please cite the MDRS benchmark paper. Citation details will be added here once the paper is public.