A minimum implementation of pose warper https://arxiv.org/pdf/1906.04016.pdf
This repository is for research use on the problem of keypoint estimation for videos with sparse labels.
- This repository is tested successfully with PyTorch 1.4.0
- This repository is dependent on the following repositories:
- HRNet as the baseline model.
- Deformable convolution as the warper.
- End-to-end style
Train the whole model all in one.
from models.hrnet_warping import HRNet_Warping
model = HRNet_Warping(...)
- Baseline-warper separated style
Due to the big size of the model, it is difficult to train it as a whole on a commercial GPU with feasible batch sizes. Therefore, you can also separate the model to two models: baseline and warper, and train them sequentially. Actually, the author of the original paper also trained it in this way.
from models.hrnet import HRNet
from models.warping import Warping
# create and train baseline model
baseline_model = HRNet(...)
...
# collect estimation results from baseline model, then fit waper model with the estimation results
warper_model = Warping(...)