Home

Awesome

3d_pose_baseline_pytorch

A PyTorch implementation of a simple baseline for 3d human pose estimation. You can check the original Tensorflow implementation written by Julieta Martinez et al.. Some codes for data processing are brought from the original version, thanks to the authors.

<!-- ![demo](./img/demo.jpg) -->

This is the code for the paper

@inproceedings{martinez_2017_3dbaseline,
  title={A simple yet effective baseline for 3d human pose estimation},
  author={Martinez, Julieta and Hossain, Rayat and Romero, Javier and Little, James J.},
  booktitle={ICCV},
  year={2017}
}

WIP

Datasets

Dependencies

Installation

  1. First, clone this repository:
    git clone --recursive https://github.com/weigq/3d_pose_baseline_pytorch.git
    
  2. Download the pre-processed Human3.6M dataset in 3d joints:
    unzip human36m.zip
    rm h36m.zip
    

Usage

Data preprocess

Train

  1. Train on Human3.6M groundtruth 2d joints:

    # optional arguments, you can access more details in opt.py
    main.py [-h] [--data_dir DATA_DIR] [--exp EXP] [--ckpt CKPT]
               [--load LOAD] [--test] [--resume]
               [--action {all,All}]
               [--max_norm] [--linear_size LINEAR_SIZE]
               [--num_stage NUM_STAGE] [--use_hg] [--lr LR]
               [--lr_decay LR_DECAY] [--lr_gamma LR_GAMMA] [--epochs EPOCHS]
               [--dropout DROPOUT] [--train_batch TRAIN_BATCH]
               [--test_batch TEST_BATCH] [--job JOB] [--no_max] [--max]
               [--procrustes]
    

    train the model:

    python main.py --exp example
    

    You will get the training and testing loss curves like:

    log

  2. Train on Human3.6M 2d joints detected by stacked hourglass:

    <!-- ``` --> <!-- ``` --> <!-- You will get the training and testing loss curves like: --> <!-- ![log](./img/log_ft.png) -->

Test

  1. You can download the pretrained model on ground-truth 2d pose for a quick demo.

    python main.py --load $PATH_TO_gt_ckpt_best.pth.tar --test
    

    and you will get the results:

    direct.discuss.eat.greet.phonephotoposepurch.sitsitd.somkewaitwalkd.walkwalkTavg
    original version37.744.440.342.148.254.944.442.154.658.045.146.447.636.440.445.5
    pytorch version35.742.339.440.744.553.342.840.152.553.942.843.144.133.436.3-

License

MIT