The MI-Motion Dataset and Benchmark for 3D Multi-Person Motion Prediction

Anonymous submission
Supplementary Code arXiv

Dataset

Our MI-Motion Dataset comprises approximately 167k sequence frames involving 3~6 subjects. Each subject is represented by 20 body 3D keypoints, and performed interactions with other individuals with different interaction levels. To simulate various activity scenes, we divide the dataset into five subsets: park, street, indoor, special locations, and complex crowd.

Accessing the dataset

The pure IMU part of the MI-Motion dataset is now publicly available and can be directly downloaded from Baidu Disk (PWD: 2v41).

The synthetic part requires signing a data access agreement before use. Please follow these steps:

  1. Download and fill in the MI-Motion Data Access Agreement.
  2. Please send the completed and signed agreement to the email address specified in the agreement form.
  3. Once permission is granted, the download link will be sent to the email address provided in the agreement.

You can also download the pretrained models of all the baselines in Google Drive. More details could be found in the Project Page. About more data details and training instructions, please see README.


Benchmark

Please see our benchmark results on short-term and long-term prediction in the paper. We proivde our benchmark code and other baseline codes in supplementary material. For ultra-long-term prediction, we attach the visualization results below.


Park Indoor Street Special Locations Complex Crowd
HRI
MRT
TBIFormer
SocialTGCN
GT