OccludeNet Dataset Overview

Dataset Introduction

OccludeNet is a large-scale occlusion video dataset designed to address the shortage of occlusion data in existing action‑recognition video datasets. The dataset contains both real‑world and synthetic occluded scene videos, covering dynamic tracking occlusions, static scene occlusions, and multi‑view interactive occlusions across various natural environments.

Dataset Structure

OccludeNet is divided into four subsets:

OccludeNet‑D

• Contains training, validation, and test sets.
• Each set is further divided into sub‑folders representing different occlusion ratios (25%, 50%, 75%).
• Within each sub‑folder are video files belonging to different action categories.

OccludeNet‑S

• Contains training, validation, and test sets.
• Each set includes video files for various action categories.

OccludeNet‑I

• Contains training, validation, and test sets.
• Each set includes video files for various action categories.

OccludeNet‑M

• Contains video files for different action categories, with multiple view angles per category.

Dataset Updates

• Version 24.11 released a pre‑print of OccludeNet, available at: https://arxiv.org/abs/2411.15729.

Dataset Annotations

• Annotation files are provided, containing fields such as action category, filename, occlusion type, occlusion filename, occlusion pixel ratio, occlusion size ratio, occlusion duration, video length, frame rate, etc.

Citation

@misc{zhou2024occludenetcausaljourneymixedview,
      title={OccludeNet: A Causal Journey into Mixed-View Actor-Centric Video Action Recognition under Occlusions},
      author={Guanyu Zhou and Wenxuan Liu and Wenxin Huang and Xuemei Jia and Xian Zhong and Chia-Wen Lin},
      year={2024},
      eprint={2411.15729},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2411.15729}
}