Datasetpublic

Khat865/ANUBIS-Skeleton

ANUBIS Dataset ANUBIS Dataset Large-Scale Skeleton-Based Action Recognition A comprehensive multi-view skeleton action dataset for challenging real-world scenarios 📄 Paper • 🌐 Project Website • 📊 Download Dataset • 💻 Benchmark Code 📝 Overview ANUBIS is a large-scale skeleton-based action recognition dataset designed to address critical gaps in existing benchmarks. The dataset features 102 action categories collected from 80 participants… See the full description on the dataset page: https://huggingface.co/datasets/Khat865/ANUBIS-Skeleton.

sourceHugging Facecc-by-nd-4.0updated 4mo agoView on Hugging Face

3likes148downloads

Dataset Card

ANUBIS Dataset

<h1 align="center"> ANUBIS Dataset <br/> Large-Scale Skeleton-Based Action Recognition </h1>

<p align="center">A comprehensive multi-view skeleton action dataset for challenging real-world scenarios</p>

<p align="center"> <a href="https://arxiv.org/abs/2205.02071">📄 Paper</a> • <a href="https://yliu1082.github.io/ANUBIS/">🌐 Project Website</a> • <a href="https://huggingface.co/datasets/Khat865/Anubis-skeleton">📊 Download Dataset</a> • <a href="https://github.com/khat865/ANUBIS-Sourcecode">💻 Benchmark Code</a> </p>

📝 Overview

ANUBIS is a large-scale skeleton-based action recognition dataset designed to address critical gaps in existing benchmarks. The dataset features 102 action categories collected from 80 participants across 80 viewpoints, generating 66,232 skeleton clips with comprehensive multi-modal data (RGB, Depth, 3D Skeleton).

Key Innovations

—🎯 Multi-View Coverage: 80 viewpoints including challenging back-view perspectives often missing in existing datasets
—👥 Multi-Person Interactions: Complex social interactions including collaborative and aggressive behaviors
—🔍 Fine-Grained Actions: Detailed hand/object manipulations and contemporary social behaviors
—⚔️ Security-Critical Actions: Violent and aggressive actions essential for surveillance applications
—🦠 Modern Social Behaviors: Pandemic-era gestures and social distancing protocols

📊 Dataset Statistics

Attribute	Value
Action Categories	102
Participants	80
Total Clips	66,232
Viewpoints	80
Modalities	RGB, Depth, 3D Skeleton
Skeleton Joints	32 per person
Frame Length	300 frames

Action Category Distribution

—Independent Actions: 45 classes (44.1%) - Single-person behaviors
—Aggressive Actions: 40 classes (39.2%) - Security-critical interactions
—Social Interactive Actions: 15 classes (14.7%) - Collaborative behaviors
—Other Actions: 2 classes (2.0%) - Spatial positioning changes

If you want to download this dataset to your local machine, we recommend that you download the anubis.zip file.

🎯 Key Challenges

1.Viewpoint Variation: Back-view perspectives with joint occlusions
2.Fine-Grained Recognition: Subtle hand gestures and object manipulations
3.Multi-Person Dynamics: Complex interpersonal interactions
4.Action Similarity: Semantically similar actions with different contexts
5.Motion Complexity: Varying temporal scales and movement patterns

📚 Citation

If you use the ANUBIS dataset in your research, please cite:

bibtex

@article{LIU2026114140,
title = {Representation-centric survey of supervised skeletal action recognition and the new benchmark},
journal = {Pattern Recognition},
pages = {114140},
year = {2026},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2026.114140},
url = {https://www.sciencedirect.com/science/article/pii/S0031320326011052},
author = {Yang Liu and Jiyao Yang and Madhawa Perera and Pan Ji and Dongwoo Kim and Min Xu and Tianyang Wang and Saeed Anwar and Tom Gedeon and Lei Wang and Zhenyue Qin}
}

📜 License

This dataset is released under [license: cc-by-nd-4.0]. Please ensure compliance with ethical guidelines when using data containing human subjects.

🙏 Acknowledgments

We thank all participants who contributed to data collection and the research community for their valuable feedback in developing this benchmark dataset.