maelic/PSG-coco-format

PSG — Panoptic Scene Graph (COCO format)

This dataset is a reformatted version of the Panoptic Scene Graph (PSG) benchmark (Yang et al., NeurIPS 2022) in standard COCO-JSON format, ready for use with object detection and scene graph generation pipelines.

It was produced as part of the SGG-Benchmark framework and used to train the models described in the REACT paper (Neau et al., BMVC 2025).

/!\ Disclaimer: this dataset does NOT contain original segmentation masks, but only bounding boxes and category labels. Thus this is NOT a panoptic dataset, but rather a scene graph dataset that can only be used to train bounding-box-based SGG models. The original PSG dataset can be downloaded from the PSG project page.

---

Annotation overview

Each image comes with:

• Object bounding boxes — 133 COCO object categories.
• Scene-graph relations — 56 predicate categories connecting pairs of objects as directed (subject, predicate, object) triplets.

Four random validation images with bounding boxes (coloured by category) and relation arrows (yellow, labelled with the predicate name).

---

Dataset statistics

Split	Images	Object annotations	Relations
train	45 564	494 213	254 214
val	1 000	19 039	7 458
test	2 186	24 910	13 705

---

Object categories (133)

Standard 133-class COCO panoptic vocabulary: person, bicycle, car, motorcycle, airplane, bus, train, truck, boat, traffic light, … (full list embedded in dataset_info.description).

Predicate categories (56)

over · in front of · beside · on · in · attached to · hanging from · on back of · falling off · going down · painted on · walking on · running on · crossing · standing on · lying on · sitting on · flying over · jumping over · jumping from · wearing · holding · carrying · looking at · guiding · kissing · eating · drinking · feeding · biting · catching · picking · playing with · chasing · climbing · cleaning · playing · touching · pushing · pulling · opening · cooking · talking to · throwing · slicing · driving · riding · parked on · driving on · about to hit · kicking · swinging · entering · exiting · enclosing · leaning on

---

Dataset structure

DatasetDict({
    train: Dataset({
        features: ['image', 'image_id', 'width', 'height', 'file_name',
                   'objects', 'relations'],
        num_rows: 45564
    }),
    val: Dataset({
        features: ['image', 'image_id', 'width', 'height', 'file_name',
                   'objects', 'relations'],
        num_rows: 1000
    }),
    test: Dataset({
        features: ['image', 'image_id', 'width', 'height', 'file_name',
                   'objects', 'relations'],
        num_rows: 2186
    }),
})

Each row contains:

Field	Type	Description
image	Image	PIL image
image_id	int	Original COCO image id
width / height	int	Image dimensions
file_name	str	Original filename
objects	List[dict]	{id, category_id, bbox (xywh), area, iscrowd, segmentation}
relations	List[dict]	{id, subject_id, object_id, predicate_id} — ids refer to objects[*].id

---

Usage

from datasets import load_dataset
import json

ds = load_dataset("maelic/PSG-coco-format")

# Recover label maps from the embedded metadata
meta = json.loads(ds["train"].info.description)
cat_id2name  = {c["id"]: c["name"] for c in meta["categories"]}
pred_id2name = {c["id"]: c["name"] for c in meta["rel_categories"]}

sample = ds["train"][0]
image  = sample["image"]          # PIL Image
for obj in sample["objects"]:
    print(cat_id2name[obj["category_id"]], obj["bbox"])
for rel in sample["relations"]:
    print(rel["subject_id"], "--", pred_id2name[rel["predicate_id"]], "->", rel["object_id"])

---

Citation

If you use this dataset, please cite the original PSG paper:

@inproceedings{yang2022panoptic,
  title        = {Panoptic scene graph generation},
  author       = {Yang, Jingkang and Ang, Yi Zhe and Guo, Zujin and Zhou, Kaiyang
                  and Zhang, Wayne and Liu, Ziwei},
  booktitle    = {European conference on computer vision},
  pages        = {178--196},
  year         = {2022},
  organization = {Springer},
}

And the REACT paper if you use the SGG-Benchmark models:

@inproceedings{Neau_2025_BMVC,
  author    = {Ma\"elic Neau and Paulo Eduardo Santos and Anne-Gwenn Bosser
               and Akihiro Sugimoto and Cedric Buche},
  title     = {REACT: Real-time Efficiency and Accuracy Compromise for Tradeoffs
               in Scene Graph Generation},
  booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025,
               Sheffield, UK, November 24-27, 2025},
  publisher = {BMVA},
  year      = {2025},
  url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_239/paper.pdf},
}

---

License

This dataset inherits the MIT license of the original PSG benchmark. See the MIT License for details.