3dpose_gan_pytorch
Pytorch implementation of Unsupervised Adversarial Learning of 3D Human Pose from 2D Joint Locations
Unsupervised Adversarial Learning of 3D Human Pose from 2D Joint Locations
This is a modified version of the project Unsupervised Adversarial Learning of 3D Human Pose from 2D Joint Locations
The Chainer code has been substituted with an equivalent version of the PyTorch code. NOTE: I was unable to compare the results with Chainer implementation due to HW/SW issues.
Run Inference for demo (with openpose)
- Download openpose pretrained model
- openpose_pose_coco.prototxt
- pose_iter_440000.caffemodel
- 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 or python bin/train.py --gpu 0 --mode unsupervised --dataset h36m --use_heuristic_loss --use_bn
-
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
python bin/eval.py
python bin/eval.py --gpu 0 --mode unsupervised --activate_func relu --use_bn -d COCO (error with --use_bn when not trained with bn)