All the raw data we used to train and evaluate the BiRNN model we presented in the paper can be downloaded here. We also made the trained BiRNN model, the results of SOP, SIP and DIP, and the intermediate files used for training public for easy reproduction. More specifically, these files are:

  • The large-scale Synthetic Data generated via SMPL body model from AMASS, the collection of archival MoCap data.
  • The real DIP-IMU dataset we captured ourself via 17 XSens IMUs. The proxy for Ground-Truth pose parameter is obtained via SIP.
  • The per-frame normalized Synthetic Data, which is directly trainable with the code we released.
  • The per-frame normalized DIP-IMU data, for fine-tuning and testing. Compatible with the code we released.
  • The results of SOP, SIP and DIP on DIP-IMU.

Please login or sign up to be able to access DIP data.


DIP model

Pre-trained models for offline and online evaluation as reported in the paper. Please refer to the README in the code repository for more details how to use it.