People-Ojbect Interaction Dataset

Introduction

The people-object interaction dataset comprises 38 series of 30-view multi-person or single-person RGB-D video sequences, complemented by corresponding camera parameters, foreground masks, SMPL models and some point clouds, mesh files. Each video sequence boasts a 4K resolution, 25 FPS, and a duration of 1~19 seconds. All 30 views are captured using Kinect Azure devices in a uniformly surrounding scene. Github repository of our dataset is https://github.com/sjtu-medialab/People-Ojbect-Interaction-Dataset.

Video Sequences Collection and Post Processing

We designed the collecting environment and system as shown below.

Details of collected video sequences are show below.

For the post processing, we use BackgroundMattingV2 to extract foreground masks, methods of Zhou et al. and Zhou et al. to extract point clouds and mesh files, MMHuman3D to extract SMPL models.

Dataset Download

The name of subfolders in Meshes and Point Clouds are defined by:

4D_seriesname_frameindex,

where the seriesname corresponds to the name of subfolders in RGB-D-Mask Sequences, the frameindex are the frame index the meshes, and point clouds are constructed for.

Camera parameters

Camera parameters can be found in https://github.com/sjtu-medialab/People-Ojbect-Interaction-Dataset. Camera parameters are provided in the intrinsic.txt and extrinsic.txt. Extrinsics of our cameras are in the format of camera coordinate system to world coordinate system.

Citation

If you find our dataset useful in your research and you want to cite our dataset, please refer to https://github.com/sjtu-medialab/People-Ojbect-Interaction-Dataset.