Human-Object Interaction Motion Dataset

1. This dataset contains:

  •    ●  8 daily human-object interactions
  •    ●  More than 4000 samples, about 2.4G
  •    ●  More than 100 participants’ motion 
  •    ●  3D skeleton sequences, recorded by inertial motion capture suit in 240fps
  •    ●  Sample code for visualizing 3D skeleton sequences

2. Eight interaction motion and labels setting

Object properties


Labels meaning

Lifting a box for weight (A1) 



Moving a bowl for fragility (A2) 


Empty, Full of rice, Full of water

Walking for path width (A3) 


Narrow, Medium, Wide

Fishing for length of rod (A4) 


Short, Medium, Long

Pouring for type of liquid (A5) 


Water, Shampoo, Rice

Bending for stiffness (A6) 


Weak, Medium, Strong

Sitting for softness of chair (A7) 


Hard stool, Soft stool, Horizontal stool, Yoga ball

Drinking for water amount (A8) 


Empty, Half, Full

3. The diagram of the interactive action is as follows:

4. Data sample naming rules are as follows:AaaaSsssTtttPppp(e.g. A001S002T003P001)

  •    ●  A(Action):   Represents 8 different interaction motions.
  •    ●  S(Subject): Represents participants
  •    ●  T(Task):      Represents sub-tasks. In lifting and moving motion, it represents move the box or bowl to 4 platforms for different heights.
  •                              In walking it represents “go forward” or “go back”.
  •    ●  P(Property):  Object property label.

The datasets are released for academic research only, and are free to researchers from educational or research institutes for non-commercial purposes.

Downloads (faster for people in China)

Downloads (faster for people in other places)