I uploaded code and checkpoint for the pretrained model (trained on the Human 3.6m dataset). Formatting of the input and usage is provided comments found in src/core/pretrained_model.
Code was run with the following dependencies:
json-tricks==3.15.5
kornia==0.5.2
natsort==7.1.0
numpy==1.19.5
opencv-python==4.5.1.48
PyYAML==5.4.1
scipy==1.6.0
tensorboard==2.4.1
torch==1.8.1
torchvision==0.9.1
Run pip install -r requirements.txt to install packages.
Since we cannot provide the datasets, you have to implement your own Dataset class in src/core/utils/dataset which should return two rgb images with shape [3, height, width] containing the same person and background, but in different poses.
Template and anchor points are saved and loaded as json files.
template.json is a dict of length num_parts which specifies the location and size/shape and angle of the body parts represented by 2d gaussians (x: columns, y:rows).
anchor_points.json is also a dict of the same length specifying anchor points for each body part (x: rows, y:columns).
Check the two provided json files for details.
The order of body parts should be the same in both files.
If you use your own template and anchor points, make sure to change src/core/utils/helper/load_anchor_points and src/core/utils/losses/compute_anchor_loss according to your defintions.
For training, adjust the config file in configs as well as train.py with the paths to your data and run python train.py --config path/to/your_config_file
We'll provide checkpoints and inference code for the model trained on human36m and our infant dataset soon, stay tuned for that.
If you find this code useful for your own project/paper, please cite our work:
@inproceedings{LSchmidtke2021,
author = {Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz},
title = {Unsupervised Human Pose Estimation through Transforming Shape Templates},
booktitle = {Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
year = {2021},
}