Unsupervised Adversarial Learning of 3D Human Pose from 2D Joint Locations

This is the authors' implementation of Unsupervised Adversarial Learning of 3D Human Pose from 2D Joint Locations

Run Inference for demo (with openpose)

1. Download openpose pretrained model
   • openpose_pose_coco.prototxt
     • https://github.com/opencv/opencv_extra/blob/3.4.1/testdata/dnn/openpose_pose_coco.prototxt
   • pose_iter_440000.caffemodel
     • http://posefs1.perception.cs.cmu.edu/OpenPose/models/pose/coco/pose_iter_440000.caffemodel
2. Run Inference
   • python bin/demo.py sample/image.png --lift_model sample/gen_epoch_500.npz --model2d pose_iter_440000.caffemodel --proto2d openpose_pose_coco.prototxt
   • Need OpenCV >= 3.4
     • < 3.3 results extreamly wrong estimation

Dependencies(Recommended versions)

• Python 3.6.5
• Cupy 4.0.0
• Chainer 4.0.0
• OpenCV 3.4 (when showing results)
• git-lfs
  • to download pre-trained model
  • or you can download pre-trained model directory from https://github.com/DwangoMediaVillage/3dpose_gan/blob/master/sample/gen_epoch_500.npz?raw=true

Training

Human3.6M dataset

• Unsupervised learning of 3D points from ground truth 2D points

  python bin/train.py --gpu 0 --mode unsupervised --dataset h36m --use_heuristic_loss
  

• Unsupervised learning of 3D points from detected 2D points by Stacked Hourglass

  TBA

• Supervised learning of 3D points from ground truth 2D points

  python bin/train.py --gpu 0 --mode supervised --activate_func relu --use_bn
  

MPII dataset

TBA

MPI-INF-3DHP dataset

TBA

Evaluation

python bin/eval.py results/hoge/gen_epoch_*.npz