Awesome

Boosting Monocular 3D Human Pose Estimation with Part Aware Attention

Environment settings

The codebase is tested under the following environment settings:

• cuda: 11.0
• python: 3.8.10
• pytorch: 1.7.1
• torchvision: 0.8.2
• scikit-image: 0.18.2
• einops: 0.3.0
• timm: 0.4.9
• pyyaml: 6.0
• easydict: 1.9
• opencv-python: 4.5.2.54

Prepare the dataset

To perform the evaluation on the Human3.6M dataset, you should:

1. Download data.zip from https://cloud.tsinghua.edu.cn/f/b102a975ff8d4ae1a4c1/?dl=1.
2. Extract the file.
3. Put the extracted files into ./data/ directory.

After doing so, the file structure should be as follows:

./data
    data_2d_h36m_cpn_ft_h36m_dbb.npz
    data_2d_h36m_gt.npz
    data_3d_h36m.npz

Download the checkpoints

The trained checkpoints can be downloaded from https://cloud.tsinghua.edu.cn/d/fae76890154a45a99b31/. After downloaded, the checkpoints should be put into the ./checkpoint/ directory and the file structure of ./checkpoint/ should be as follows.

./checkpoint
    cpn_f81.bin
    cpn_f243.bin
    gt_f81.bin
    gt_f243.bin

Evaluate

To conduct evaluation using the CPN inputs, you can run the following commands:

CUDA_VISIBLE_DEVICES=0 python eval.py -c ./exp/exp_cpn_f81.yaml # using 81 frames as input

CUDA_VISIBLE_DEVICES=0 python eval.py -c ./exp/exp_cpn_f243.yaml # using 243 frames as input

Similarly, to evaluate using the GT inputs, you can run the following commands:

CUDA_VISIBLE_DEVICES=0 python eval.py -c ./exp/exp_gt_f81.yaml # using 81 frames as input

CUDA_VISIBLE_DEVICES=0 python eval.py -c ./exp/exp_gt_f243.yaml # using 243 frames as input

Acknowledgement

Part of the code is borrowed from Poseformer and VideoPose3D. We thank the authors for releasing their codes.