People-Object Interaction Dataset

People‑Object Interaction Dataset Overview

Content

• Video Sequences: 38 series with 30‑view RGB‑D videos, 4K resolution, 25 FPS, 1–19 s duration.
• Auxiliary Data: Camera parameters, foreground masks, SMPL models, point clouds, and mesh files.

Sequence Details

• Empty Scene: 1 sequence, 1 s.
• Camera Calibration: 1 sequence, 8 s.
• Single‑Person Interaction: 23 sequences, 2–19 s, covering activities such as reading, sitting, opening an umbrella, etc.
• Two‑Person Interaction: 11 sequences, 2–14 s, involving cooperative tasks like moving a table or sweeping.
• Three‑Person Interaction: 2 sequences, 2–5 s, mainly group photo scenarios.

Processing

• Foreground Masks: Extracted using BackgroundMattingV2.
• Point Cloud & Mesh: Generated following methods from Zhou et al. (IEEE 2022) and Zhou et al. (IBCV 2023).
• SMPL Models: Obtained via MMHuman3D.

Download

• Accessible via BaiduNetdisk or the Medialab portal.

Camera Parameters

• Files intrinsic.txt and extrinsic.txt provide intrinsic and extrinsic calibration data. Extrinsics map camera coordinates to world coordinates.

Citation

• Accepted at ICIP‑2024.

@inproceedings{POID,
  title={A New People‑Object Interaction Dataset and NVS Benchmarks},
  author={Guo, Shuai and Zhong, Houqiang and Wang, Qiuwen and Chen, Ziyu and Gao, Yijie and Yuan, Jiajing and Zhang, Chenyu and Xie, Rong and Song, Li},
  booktitle={2024 IEEE International Conference on Image Processing (ICIP)},
  year={2024},
  organization={IEEE}
}