Panorama-grounding/COCONut-PanCap-Recaptioned

COCONut-PanCap Re-captioned (train2017)

Regenerated captions for the train2017 split of COCONut-PanCap (paper, dataset).

The released captions contain phrases bound to the wrong segment, references to objects that are not in the image, redundant restatements, and untagged summary sentences that carry no grounding. We regenerated every caption and repaired the remaining errors, keeping the file layout of the original release.

Only the caption text changed. Mask ids, categories and the images are untouched, so this drops into an existing COCONut-PanCap pipeline with no code changes. It covers 118,062 images.

See also our human-annotated benchmark for panoptic grounded captioning, PanoCaps.

Images

Images and masks are not included in this repository:

	Source
Images	COCO train2017
Panoptic masks	xdeng77/coconut_s

Captions cover every train image that has at least one mask segment. Images with no segments, and images outside the COCONut-PanCap train inventory, have no caption here.

Download

hf download Panorama-grounding/COCONut-PanCap-Recaptioned --repo-type dataset \
    --local-dir /path/to/data/COCONut_PanCap
tar -xf /path/to/data/COCONut_PanCap/caption_train2017.tar -C /path/to/data/COCONut_PanCap

The panoptic masks are published as PNGs on Kaggle and as parquet on xdeng77/coconut_s; unpack the parquet with:

import glob, os, pyarrow.parquet as pq
out = '/path/to/data/COCONut_PanCap/coconut_s/panoptic'
os.makedirs(out, exist_ok=True)
for f in sorted(glob.glob('/path/to/coconut_s/data/*.parquet')):
    for b in pq.ParquetFile(f).iter_batches(batch_size=512, columns=['mask', 'segments_info']):
        for mask, info in zip(b.column('mask').to_pylist(), b.column('segments_info').to_pylist()):
            open(f"{out}/{info['file_name']}", 'wb').write(mask['bytes'])

Formats

The same captions are published two ways:

• data/ — parquet, for browsing in the Dataset Viewer and for quick loading.
• caption_train2017.tar — one .txt per image, identical in layout and naming to the original release. Use this as a drop-in replacement.

coconut_s_panoptic_train2017.json accompanies them: the COCO-panoptic annotation file a loader needs to resolve mask ids to segments, with images[*] and annotations[*].segments_info giving each segment's id, category_id and isthing. COCONut publishes its masks only as parquet, so this file is not otherwise available.

Caption records

data/train-*.parquet:

Field	Description
image_id	COCO image id, e.g. 24769.
file_name	Image file name, e.g. 000000024769.jpg.
caption	Caption text with inline <ids: description> grounding tags.
label_matched	List of {"mask_ids": [...], "txt_desc": "..."}, one per grounded phrase.

mask_ids are the panoptic segment ids used by COCONut, matching segments_info[*].id.

Caption text

Plain text with inline <ids: description> tags, where ids are one or more comma-separated panoptic segment ids. The figure above shows a caption before and after re-captioning.

Re-captioning

Captions were regenerated with Qwen3-VL-235B-A22B-Instruct, prompting the model to describe each image while binding every mentioned object to its panoptic mask id. A second stage then corrected the remaining errors: untagged opening and closing summaries were removed, counts in grouped tags were fixed, phrases bound to the wrong same-category segment were re-pointed to the correct one, and phrases asserting a real person or animal where the mask is a depiction or a reflection were rewritten to say so. Redundant text was trimmed where a description re-identified an object by whoever holds it, repeated an action its own sentence already states, or duplicated a phrase that appears verbatim in the surrounding prose.

Every caption parses cleanly and every mask id it references resolves to a real segment. No correction ever altered an image's set of ids, so grounding coverage is identical to the original release. A small fraction of segments are referenced by no phrase at all, almost always individual items inside large repeated groups such as produce stalls or crowds, where the caption grounds the group rather than each instance.

Residual issues, stated honestly: binding errors remain possible in images holding several instances of the same category, which is where review was concentrated; images outside that slice were not searched for phrases describing depictions rather than real objects, so their rate is unmeasured; and in large crowds the per-instance attributes are unreliable even though the grounding itself is correct.

License

This dataset builds upon COCONut-PanCap and MS COCO; the images and panoptic masks remain under their original licenses. The re-captioned annotations are released for research and non-commercial use only.

Citation

If you use these captions, please cite our paper and the original COCONut-PanCap dataset:

@article{pieri2026panorama,
  title   = {PANORAMA: Panoptic Grounded Captioning via Mask Proposal Selection},
  author  = {Pieri, Sara and Kazakos, Evangelos and Chen, Shizhe and Sivic, Josef and Schmid, Cordelia},
  journal = {arXiv preprint arXiv:2609.19143},
  year    = {2026}
}

@inproceedings{deng2025coconutpancap,
  title     = {{COCONut-PanCap}: Joint Panoptic Segmentation and Grounded Captions for Fine-Grained Understanding and Generation},
  author    = {Deng, Xueqing and Yu, Qihang and Athar, Ali and Yang, Chenglin and Yang, Linjie and Jin, Xiaojie and Shen, Xiaohui and Chen, Liang-Chieh},
  booktitle = {Advances in Neural Information Processing Systems Datasets and Benchmarks Track},
  year      = {2025}
}