Panorama-grounding/MRSeg-Referring-Expressions

MRSeg Referring Expressions (single-turn)

Single-turn referring-expression segmentation samples derived from the multi-round MR-Seg data of SegLLM (paper).

SegLLM's original conversations provide the target regions as input to later rounds, which makes them unusable for models that receive only an image and text. We used Qwen3-VL-32B-Instruct to keep only the samples whose target can be identified from the image and expression alone, rewriting each into a self-contained referring expression. The segmentation masks are resolved from the source datasets' annotations and stored inline as COCO RLE, so no external annotation files are needed — only the images.

The release contains 101,349 samples spanning parts, stuff regions, attributes, and relationally specified regions. This is the multi-granularity bucket of PANORAMA's training mixture.

See also our human-annotated benchmark for panoptic grounded captioning, PanoCaps, and our re-captioned COCONut-PanCap.

What is in this repository

Path	Contents
data/train-*.parquet	one row per sample, mask inline as COCO RLE
mrseg_accepted.tar	the same samples as per-source .jsonl shards, as read by our loader

Download

hf download Panorama-grounding/MRSeg-Referring-Expressions --repo-type dataset \
    --local-dir /path/to/data/MRSeg-Referring-Expressions
mkdir -p /path/to/data/mrseg_extract && tar -xf /path/to/data/MRSeg-Referring-Expressions/mrseg_accepted.tar \
    -C /path/to/data/mrseg_extract   # accepted/*.jsonl

Images

Images are not included. Samples reference them by a relative path (img_rel) under four public sources:

img_rel prefix	Dataset	Download
coco/	COCO train2017	cocodataset.org
ade20k/	ADE20K	MIT Scene Parsing
pascal/	PASCAL VOC 2010	host.robots.ox.ac.uk
VG_100K/, VG_100K_2/	Visual Genome	visualgenome.org

LVIS, PACO and COCO-Stuff annotate COCO images, so they require no additional image download.

Fields

Field	Description
sid	Unique sample id, e.g. ade20k_016857_00.
source	Source bucket: mr_lvis, mr_paco, mr_pascal, mr_vg, ade20k, cocostuff or attributes.
img_rel	Image path relative to the source dataset root.
expression	Self-contained referring expression, e.g. "the stone stairs in the colonnade".
instruction	A natural-language segmentation instruction wrapping the expression (instr in the jsonl).
mask	COCO RLE (size, counts) in the jsonl; flattened to mask_size / mask_counts in the parquet.
model, prompt_ver	Extraction provenance (jsonl only).

License

This dataset is derived from SegLLM's MR-Seg data and from the annotations of LVIS, PACO, Visual Genome, COCO-Stuff, ADE20K and PASCAL-Part; it follows the licenses of these source datasets. For research and non-commercial use only.

Citation

If you use these expressions, please cite our paper and the original SegLLM work:

@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{wang2025segllm,
  title     = {{SegLLM}: Multi-round Reasoning Segmentation with Large Language Models},
  author    = {Wang, XuDong and Zhang, Shaolun and Li, Shufan and Kallidromitis, Konstantinos and Li, Kehan and Kato, Yusuke and Kozuka, Kazuki and Darrell, Trevor},
  booktitle = {International Conference on Learning Representations},
  year      = {2025}
}