RAWDet-7: object detection

RAWDet-7 is a multi-scenario benchmark for object detection on full-precision RAW images and their corresponding high-resolution sRGB images. It consolidates PASCAL RAW, RAOD, Zurich RAW, and NOD (Nikon and Sony) under seven standardized categories: car, truck, tram, person, bicycle, motorcycle, and bus.

This is the complete paper release: 32,552 RAW/sRGB pairs, 251,168 retained detections, complete COCO and per-image annotations, corrected split manifests, and portable metadata. MultiRAW was a comparison dataset in the paper and is intentionally not included.

Dataset statistics

Source	Train	Test	Sensor / precision	Scenario
PASCAL RAW	2,128	2,130	Nikon D3200, 12-bit	Day
Zurich RAW	104	25	Huawei P20 / Sony IMX380, 10-bit effective precision	Day
NOD Nikon	3,373	843	Nikon D750, 14-bit	Night
NOD Sony	3,309	828	Sony RX100 VII, 14-bit	Night
RAOD	15,950	3,862	Sony IMX490, packed 24-bit	Day and night
Total	24,864	7,688

The paper calls the held-out split test; the extracted directory and paper code call it val. Hugging Face exposes it as test while all asset paths retain val/ for compatibility.

Category IDs are stable: 1 car, 2 truck, 3 tram, 4 person, 5 bicycle, 6 motorcycle, and 7 bus. The canonical COCO files contain 201,602 train and 49,566 test annotations after the Grounded-DINO confidence threshold of 0.8.

Why the images are archived

The code-compatible tree contains 15,950 RAOD training files in each of three flat directories, exceeding the Hugging Face Hub limit of 10,000 entries per folder. The original verified 69-volume ZIP is therefore published intact. One extraction command recreates the exact flat paths expected by the paper code. COCO annotations and lightweight Parquet indices remain directly downloadable.

RawDet-7 Hub repository/
├── annotations_coco/                 # directly usable, 12 JSON files
├── archives/                         # RAWDet-7.z01 ... z68 + RAWDet-7.zip
├── data/{train,test}-00000-of-00001.parquet
├── examples/rawdet7_hf_dataset.py
├── metadata/
│   ├── archive_inventory.csv
│   ├── extracted_file_manifest.tsv
│   └── release_summary.json
├── release_overlay/portable_release_metadata.tar.gz
└── scripts/extract_rawdet7.py

The overlay makes every per-image JSON path portable, adds the one packaged PASCAL record that was missing despite being present in COCO, replaces four stale legacy text-list entries with exact COCO-derived lists, and adds the indices and tested config example. It is applied only to a newly extracted destination; source archives are never modified.

Download and extract the complete dataset

from huggingface_hub import snapshot_download

repo = snapshot_download(
    repo_id="shashankskagnihotri/RawDet-7",
    repo_type="dataset",
    local_dir="/path/to/downloaded/RawDet-7-Hub",
)
print(repo)

Install 7-Zip, then extract and validate all required paths and sizes:

python /path/to/downloaded/RawDet-7-Hub/scripts/extract_rawdet7.py \
  --repo-root /path/to/downloaded/RawDet-7-Hub \
  --output-dir /path/to/datasets

The result is /path/to/datasets/RawDet-7 with:

RawDet-7/
├── annotations_coco/
├── annotations_normal_json/{train,val}/{PASCAL_RAW,RAOD,NIKON,SONY,ZURICH}/
├── combined_raw/{train,val}/{PASCAL_RAW,RAOD,NIKON,SONY,ZURICH}/
├── combined_sRGB/{train,val}/{PASCAL_RAW,RAOD,NIKON,SONY,ZURICH}/
├── data/
└── txt_files/

For annotations and indices only, avoid downloading the archives:

snapshot_download(
    repo_id="shashankskagnihotri/RawDet-7",
    repo_type="dataset",
    local_dir="/path/to/RawDet-7-annotations",
    allow_patterns=["annotations_coco/**", "data/**", "metadata/**", "README.md"],
)

The Parquet index is also available through:

from datasets import load_dataset

dataset = load_dataset("shashankskagnihotri/RawDet-7")
print(dataset["train"][0])

Use with the paper MMDetection code

If the extracted dataset is /path/to/datasets/RawDet-7, configure the snapshot itself as the data root:

data_root = "/path/to/datasets/RawDet-7"

train_dataloader = dict(dataset=dict(
    type="RawDetDataset",
    data_root=data_root,
    ann_file="annotations_coco/train_combined.json",
    data_prefix=dict(img=""),
))

val_dataloader = dict(dataset=dict(
    type="RawDetDataset",
    data_root=data_root,
    ann_file="annotations_coco/val_combined.json",
    data_prefix=dict(img=""),
))

Use the same data_root in the train/test LoadImageFromFile pipeline entries. examples/rawdet7_hf_dataset.py is a copy-ready override for the project’s Grounding-DINO config. Install rawpy before RAW-mode runs; it is imported by the historical project but absent from its requirements.

RAW loading follows the paper implementation:

• PASCAL RAW (.nef), NOD Nikon (.NEF), and NOD Sony (.ARW) use rawpy.
• RAOD stores tightly packed little-endian 24-bit values (three bytes per pixel), reshaped to 1856 × 2880.
• Zurich RAW uses .png with 10-bit effective precision.

Quantized 4-, 6-, and 8-bit variants are produced at runtime; they are not duplicated in this release.

Zurich RAW PNGs are 3968 × 2976, while their matching sRGB images and COCO coordinate frame are 5120 × 3840. The historical loader does not rescale those boxes; add coordinate-aware resizing for Zurich-only RAW benchmarking.

Provenance and limitations

The annotations were generated with Grounded-DINO 1.5 Pro using car . truck . tram . person . bicycle . motorcycle . bus ., filtered at confidence 0.8, and validated on the corresponding sRGB images. The release preserves RLE segmentations, confidence scores, and original box coordinates. A small number of boxes extend a few pixels beyond image boundaries because of floating-point edge noise; consumers may clamp at load time.

The packaged COCO files contain 251,168 annotations, 30 more truck instances than the total implied by Figure 13 of the manuscript. The released annotation files are authoritative; this discrepancy is documented rather than silently altering labels.

RAWDet-7 combines images originating from PASCAL RAW, RAOD, Zurich RAW, and NOD. Cite RAWDet-7 and the relevant upstream source datasets when using a subset.

License and citation

The official University of Mannheim dataset host distributes RAWDet-7 under the Creative Commons Attribution 4.0 International license. Source-dataset provenance remains applicable; users are responsible for relevant upstream terms.

@article{fatima2026rawdet7,
  title   = {{RAWDet-7}: A Multi-Scenario Benchmark for Object Detection and Description on Quantized {RAW} Images},
  author  = {Fatima, Mishal and Agnihotri, Shashank and Gandikota, Kanchana Vaishnavi and Moeller, Michael and Keuper, Margret},
  journal = {arXiv preprint arXiv:2602.03760},
  year    = {2026}
}

Paper: OpenReview · arXiv · Code: Mishalfatima/RawDet-7