MDVD-108K

MDVD-108K is a VisDrone-based benchmark for all-in-one UAV image restoration and detection robustness. It contains 108,000 paired synthetic samples across 12 single and compound degradations, plus 500 real degraded images.

The release uses capacity-bounded Parquet shards. It avoids hundreds of thousands of repository files while supporting the Hugging Face Dataset Viewer, streaming, filtering, and selective downloads. Original JPEG bytes and VisDrone annotations are preserved without image recompression. The previous sample-aligned WebDataset layout remains available under the webdataset-v2 tag.

Dataset Structure

split	paired samples	real samples
train	86,400	0
validation	10,800	0
test	10,800	0
real	0	500

Each row provides:

• input: degraded 512 x 512 image
• target: clean 512 x 512 reference; null for the real split
• annotation: original VisDrone detection annotation
• sample_id, split, degradation, source_filename, width, and height

Synthetic splits cover:

defocus, motion, fog, rain, dust, lowlight,
fog-motion, fog-lowlight, dust-motion, dust-lowlight,
rain-motion, rain-lowlight

Stream With Datasets

from datasets import load_dataset

dataset = load_dataset("zzaiyan/MDVD-108K", split="train", streaming=True)
sample = next(iter(dataset))

degraded = sample["input"]
target = sample["target"]
degradation = sample["degradation"]

Remove streaming=True to prepare a local Arrow dataset with random access.

Download Shards

Download the full release:

hf download zzaiyan/MDVD-108K --repo-type dataset --local-dir MDVD-108K

Select a split or degradation without downloading the rest:

hf download zzaiyan/MDVD-108K --repo-type dataset \
  --include "data/test/**" --local-dir MDVD-108K

hf download zzaiyan/MDVD-108K --repo-type dataset \
  --include "data/test/fog/**" --local-dir MDVD-108K

Restore the conventional directory tree when existing code expects ordinary files:

python tools/unpack_mdvd_parquet.py MDVD-108K /path/to/MDVD-108K-extracted

This produces train|val|test/{degradation}/{input,gt,label} and real/{input,label}. MANIFEST.csv records shard sizes and sample counts, while SHA256SUMS supports integrity verification:

sha256sum --check SHA256SUMS

See ANNOTATION_FORMAT.md for label fields and the included YOLO/COCO conversion tools.