Deeppose

• [This is not an official code] Original paper is DeepPose: Human Pose Estimation via Deep Neural Networks.
• Original code we referenced: mitmul/Deeppose
• We modified the code to match the versions of the libraries below. However, MPII is the only dataset we performed. other datasets have not been performed due to url error.

Tested Environment

• Ubuntu 20.0.4 LTS
• Anaconda 22.9.0
• GPU : Nvidia RTX 3060
• GPU Setting
  • Nvidia-driver 525.105.17
  • CUDA 11.8
  • cuDNN 8.9.2
  • CuPy 7.8
• Python 3.7.5
• Chainer 7.8.0
• Numpy 1.21.5
• scikit-image 0.19.3
• OpenCV 3.4.2

Dataset preparation

 bash datasets/download.sh
 python datasets/mpii_dataset.py

Start Training

 bash shells/train_mpii.sh

Make Output

• After train, you can make output image file(.png)

• this is by evaluate_flic.py, it is possible to check the results through image file.

  python scripts/evaluate_flic.py
  --model {result-dir}/{model}.py
  --param {result-dir}/{model}.npz
  --batchsize {batchsize}
  --gpu {num} --datadir data/mpii
  --n_imgs {num} --resize {size} --seed {seed-num}
  --mode {test or tile}
  

• if you choose mode as test -> will make result image about test set

• if you choose mode as tile -> will make result image about random image in test set.

• but, you have to execute test mode first and then execute tile.