shashankskagnihotri/RawDet-7-Object-Descriptions

RAWDet-7: object-description track

This is the 500-image object-description track from RAWDet-7. It is object-level description with set-of-marks, not ordinary whole-image captioning: each annotated object is identified by a numbered black square with a white outline and a colored number, and receives its own detailed caption.

The release contains the exact 500 held-out images used by the paper, their corresponding full-precision RAW files, cleaned detection JSON, all 17 marked image variants per source image, and all structured captions and scores that were persisted by the experiment.

Statistics

• 500 unique images from the RAWDet-7 held-out (val on disk, test in the paper) split.
• 8,500 marked images: exactly 17 variants for each source image.
• 2,553 marked objects and object-level records.
• Sources: 238 RAOD, 148 PASCAL RAW, 65 NOD Nikon, and 49 NOD Sony images.
• Classes: 1,313 car, 813 person, 192 bicycle, 107 truck, 88 motorcycle, 33 bus, and 7 tram objects.
• One caption cell is blank: 2014_003525, mark 2, linear_4_bits. It is retained rather than silently imputed.

The 500-image sample is now explicit and ordered in metadata/sample_500.csv. It corresponds exactly to the first 500 rows of the filtered candidate list used by the historical run.

Contents

RawDet-7-Object-Descriptions/
├── annotations/
│   ├── detection/val/{PASCAL_RAW,RAOD,NIKON,SONY}/*.json
│   ├── per_image/*.csv
│   ├── object_descriptions_500.csv
│   ├── combined_per_image_with_sRGB_scores_corrected.csv
│   └── combined_per_image_with_sRGB_scores_corrected_I_cannot_see_the_object.csv
├── images/
│   ├── sRGB/val/{PASCAL_RAW,RAOD,NIKON,SONY}/
│   ├── raw/val/{PASCAL_RAW,RAOD,NIKON,SONY}/
│   └── overlays/  # 17 marked variants × 500 images
├── data/test-00000-of-00001.parquet
├── metadata/
│   ├── sample_500.csv
│   ├── sample_500_paths.csv
│   ├── variant_images.csv
│   ├── release_summary.json
│   └── file_manifest.tsv
├── gamma_values_combined.csv
└── gamma_values_ind.csv

The Parquet/CSV table has one row per marked object. Stable fields include source dataset, relative sRGB/RAW/annotation paths, all 17 marked-variant paths, image dimensions, mark ID, class, source xyxy bounding box, detection score, both high-resolution sRGB reference captions, candidate captions, and the experiment's BLEU/overlap/judge metrics. metadata/variant_images.csv provides a compact one-row-per-overlay index.

Gemini 2.5 Pro generated two reference descriptions from each marked high-resolution sRGB image. Candidate caption text is present for:

• RGB-Down at 4, 6, and 8 bits;
• linear RAW at 4, 6, and 8 bits;
• logarithmic RAW at 4, 6, and 8 bits;
• learned-gamma RAW at 4, 6, and 8 bits;
• log-plus-gamma RAW at 4, 6, and 8 bits.

The historical pipeline also generated a separate 50%-downsampled sRGB caption and stored its evaluation metrics, but did not persist that caption text in any structured file. The corresponding marked image is included, but the release does not invent or reconstruct its missing caption text. All 17 stored overlays are retained; large visualization PDFs and trial-analysis outputs are reproducible reports rather than dataset records and are not duplicated.

Download and inspect

from datasets import load_dataset

descriptions = load_dataset(
    "shashankskagnihotri/RawDet-7-Object-Descriptions"
)
print(descriptions["test"][0])

load_dataset retrieves the lightweight table. Download the image assets with:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="shashankskagnihotri/RawDet-7-Object-Descriptions",
    repo_type="dataset",
    local_dir="/path/to/RawDet-7-Object-Descriptions",
)

Every path in the table is relative to the snapshot root.

Use with the paper captioning code

The project expects image paths containing a literal /val/, a per-image detection annotation root, a RAW root, and gamma CSVs in the working directory. The release preserves that convention. From the downloaded dataset root:

cd /path/to/RawDet-7-Object-Descriptions

python /path/to/image_reasoning_counting/object_captioining_gemini_with_class_names_multiple_RGB_down.py \
  --paths_csv metadata/sample_500_paths.csv \
  --csv_col image_path \
  --annotations_root annotations/detection/val \
  --raw_val_root images/raw/val \
  --output_dir /path/to/new_outputs \
  --paths_per_run 50 \
  --run_id 0 \
  --judge_model_name gemini-2.5-flash-lite \
  --centre_offset 0.08 \
  --GCP_API_KEY "$GEMINI_API_KEY"

Run IDs 0 through 9 cover the ordered 500-image sample in blocks of 50. The released overlays let users consume every persisted marked input directly without reproducing the mark renderer.

The current local development copy of the generator contains an ipdb.set_trace() debugging stop after RAW-variant creation. Use the clean committed revision recorded in metadata/release_summary.json, or remove that debugging line in a personal working copy before running. This note concerns code execution only; the released data were fully validated independently.

Generation and annotation details

Detection objects are the >= 0.8 confidence RAWDet-7 annotations, in source order. Gemini 2.5 Pro received the marked image plus the ordered class list and was instructed to emit exactly one factual line per object as <number>: <caption>, describing only the indicated object. Gemini 2.5 Flash-Lite supplied the stored semantic similarity, detail-level, and detail-match judgments.

The detection JSON keeps the source boxes and RLE masks. A small number of coordinates exceed an image boundary by floating-point edge noise (maximum observed excess about 1.52 pixels); consumers may clamp boxes at load time, but source values are preserved for reproducibility.

Limitations

• Captions and judge scores are model-generated annotations, not human-authored ground truth. The paper reports separate human studies of annotation preference, caption quality, and judge correlation.
• Color/brightness changes are expected across RAW mappings; the later corrected score tables adjust judge instructions accordingly. The canonical 500 per-image caption files remain the source of caption text.
• Regenerating derived low-bit images depends on the paper code, RAW loader, and included gamma CSVs. The exact paper overlays are already present for direct use.

License

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

Citation

@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