README.md

metadata

license: mit
language:
  - en
tags:
  - research
  - scene-graphs
  - image-retrieval
  - PSG
size_categories:
  - 10K<n<100K

SCENIR - ICML-2025 - Preprocessed Dataset

This dataset is a preprocessed and refined version of the PSG dataset, specifically prepared for the research presented in our paper, "SCENIR: Visual Semantic Clarity through Unsupervised Scene Graph Retrieval". It aims to provide a ready-to-use resource for training our SCENIR model on semantic image retrieval using scene graphs.

📊 Dataset Structure

The dataset contains 3 files:

• final_train_graphs.pkl: 11054 scene graphs for the train split, with annotations, extracted sentence-transdormer embeddings, and COCO image URL/ID
• final_test_graphs.pkl: 1000 scene graphs for the test split, with annotations, extracted sentence-transdormer embeddings, and COCO image URL/ID
• ged_ground_truth.pkl: a 1000-by-1000 numpy array containing the GED (Graph Edit Distance) values for all pairs of test scene graphs (used for the evaluation)
• psg_category_embeddings.pkl: a list of several extracted embeddings of the object/relation classes of PSG, used for the computation of the ground truth GED values

🛠️ Preprocessing and Derivation

This dataset was derived from the original PSG (Panoptic Scene graph Generation) dataset by applying the following key preprocessing steps:

1. [Step 1]: Filter out PSG graphs with significantly low density.
2. [Step 2]: Remove isolated nodes
3. [Step 3]: Rename node and edge labels to remove descriptive words from PSG label vocabulary (e.g. "tree-merged" -> "tree"), to distill important information for embedding step
4. [Step 4]: Extract 768-dimensional sentence-transformer embeddings for each node/edge label to construct graph feature matrix

📜 Original Dataset & License

This dataset is a derivative work based on the PSG. We adhere to the terms of its original license.

• Original Dataset Name: Panoptic Scene Graph Generation
• Original Source: https://github.com/Jingkang50/OpenPSG
• Original License: MIT License

🤝 Citation

If you use this preprocessed dataset in your research or applications, please cite our paper and the original dataset:

@inproceedings{chaidos2025scenir,
  title={SCENIR: Visual Semantic Clarity through Unsupervised Scene Graph Retrieval},
  author={Chaidos, Nikolaos and Dimitriou, Angeliki and Lymperaiou, Maria and Stamou, Giorgos},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML)},
  year      = {2025},
  publisher = {PMLR},
  url       = {https://arxiv.org/abs/2505.15867v1},
}

@inproceedings{yang2022psg,
    author = {Yang, Jingkang and Ang, Yi Zhe and Guo, Zujin and Zhou, Kaiyang and Zhang, Wayne and Liu, Ziwei},
    title = {Panoptic Scene Graph Generation},
    booktitle = {ECCV}
    year = {2022}
}

📧 Contact

For any questions or issues regarding this dataset, please contact nchaidos@ails.ece.ntua.gr