maelic/VG150-coco-format

VG150 — Visual Genome 150 (COCO format)

This dataset is the standard VG150 split of Visual Genome (Krishna et al., 2017), the most widely used benchmark for Scene Graph Generation, reformatted in standard COCO-JSON format. VG150 contains the top 150 object categories and 50 relations from the original Visual Genome dataset, selected by frequency in the Scene Graph Generation by Iterative Message Passing paper.

This version in COCO format 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).

⚠️ Bias warning: VG150 has been heavily criticised for high class overlap and annotation biases (e.g. person / man / men / people). See VrR-VG (ICCV'19) and Neau et al. (ICCVW'23) for reference.

---

Changelog

[Update] Fixed train/val/test split leakage. A previous version of this dataset had a bug in the h5→COCO split assignment (reported in SGG-Benchmark#94): the entire train/val pool was placed in train with no validation images held out, and val was actually built from the first ~4,844 images of the test pool — i.e. val was leaked test data, and the true 5,000-image canonical validation set was silently folded into train. This has been fixed: val is now exactly the standard Neural-Motifs / Scene-Graph-Benchmark 5,000-image validation set (all with ≥1 annotated relation), disjoint from both train and test, and test is the full, untouched test pool. Object/relation annotations themselves were not affected by this bug and are unchanged content-wise. See the updated stats below. Number of images per split can be different from the original .h5 file since we are discarding images with zero object boxes (not useful for detection or SGG) and keeping images with zero relations (SGG-specific loaders already filter these out per-split at load time).

---

Annotation overview

Each image comes with:

• Object bounding boxes — 150 Visual Genome object categories.
• Scene-graph relations — 50 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	Zero-relation images kept
train	68 538	730 270	405 822	10 815
val	5 000	62 754	33 203	0 (all have ≥1 relation)
test	31 876	352 330	183 640	5 430

Images with zero object boxes are excluded entirely (not useful for detection or SGG). Images with zero relations are kept — SGG-specific loaders (sgg_benchmark/data/datasets/visual_genome.py) already filter these out per-split at load time, so relation-based training/eval effectively sees 57,723 / 5,000 / 26,446 images, matching the canonical VG150 protocol exactly, while the raw COCO annotations here stay usable for plain object detection too.

---

Object categories (150)

Top-150 Visual Genome object vocabulary used by the standard SGG split. Full list embedded in dataset_info.description.

Predicate categories (50)

and · says · belonging to · over · parked on · growing on · standing on · made of · attached to · at · in · hanging from · wears · in front of · from · for · watching · lying on · to · behind · flying in · looking at · on back of · holding · between · laying on · riding · has · across · wearing · walking on · eating · above · part of · walking in · sitting on · under · covered in · carrying · using · along · with · on · covering · of · against · playing · near · painted on · mounted on

---

Dataset structure

DatasetDict({
    train: Dataset({
        features: ['image', 'image_id', 'width', 'height', 'file_name',
                   'objects', 'relations'],
        num_rows: 68538
    }),
    val: Dataset({
        features: ['image', 'image_id', 'width', 'height', 'file_name',
                   'objects', 'relations'],
        num_rows: 5000
    }),
    test: Dataset({
        features: ['image', 'image_id', 'width', 'height', 'file_name',
                   'objects', 'relations'],
        num_rows: 31876
    }),
})

Each row contains:

Field	Type	Description
image	Image	PIL image
image_id	int	Original Visual Genome 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/VG150-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 Visual Genome:

@article{krishna2017visual,
  title={Visual genome: Connecting language and vision using crowdsourced dense image annotations},
  author={Krishna, Ranjay and Zhu, Yuke and Groth, Oliver and Johnson, Justin and Hata, Kenji and Kravitz, Joshua and Chen, Stephanie and Kalantidis, Yannis and Li, Li-Jia and Shamma, David A and others},
  journal={International journal of computer vision},
  volume={123},
  number={1},
  pages={32--73},
  year={2017},
  publisher={Springer}
}

And the original paper that established the VG150 split:

@inproceedings{xu2017scene,
  title={Scene graph generation by iterative message passing},
  author={Xu, Danfei and Zhu, Yuke and Choy, Christopher B and Fei-Fei, Li},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  pages={5410--5419},
  year={2017}
}

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

Visual Genome images and annotations are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.