A Simple Method to Boost Human Pose Estimation Accuracy by Correcting the Joint Regressor for the Human3.6m Dataset

Code repository for the paper:
A Simple Method to Boost Human Pose Estimation Accuracy by Correcting the Joint Regressor for the Human3.6m Dataset
Eric Hedlin, Helge Rhodin, Kwang Moo Yi
CRV 2022
[paper]

Using the joint regressor from the paper

The weights for the joint regressor from the paper are provided in models/retrained_J_Regressor.pt.

Training a new joint regressor

The code for training a new joint regressor is provided in scripts/optimize.py. Pseudo ground truth poses are initialized by SPIN estimates. The poses are iteratively optimized to be closer to the ground truth 3D joints while remaining realistic as according to the pose discriminator. These optimized poses are then used to supervise the new joint regressor. In theory this should be able to be applied to any dataset with corresponding ground truth 3D joints.

Citing

If you find this code useful for your research or the use data generated by our method, please consider citing the following paper:

@Inproceedings{hedlin2022refinejreg,
  Title          = {A Simple Method to Boost Human Pose Estimation Accuracy by Correcting the Joint Regressor for the Human3.6m Dataset},
  Author         = {Hedlin, Eric and Rhodin, Helge and Yi, Kwang Moo},
  Booktitle      = {CRV},
  Year           = {2022}
}