PFN

This is the release for paper "PoseFlow: A Deep Motion Representation for Understanding Human Behaviors in Videos" CVPR2018

Running

1. Download CMU Panoptic dataset, we use "Pose" subset

2. Select images and generate ground truth poseflow

• generate mat file for the dataset

run scripts/get_continuous_data.m 

you can skip this step by downloading the results form Google Drive

• sampling the data with random duration

run scripts/generate_DS_database.m  # down sampling the data

• generate poseflow ground truth

run scripts/generate_DS_poseFlow448_data.m # generate ground truth

3. Prepare caffe, we use Caffe for FlowNet2

4. Generate hmdb file before training

sh data/make-lmdb.sh

5. Training PFN

sh models/PFNST-CV/train.sh

6. Test and generate poseflow

run scripts/test_epe.m

Model download

The trained model can be downloaded from Google Drive

gitee

code has also been released in gitee

Citation

When using the code in your research work, please cite the following paper:

@inproceedings{zhang2018poseflow,
  title={PoseFlow: A Deep Motion Representation for Understanding Human Behaviors in Videos},
  author={Zhang, Dingwen and Guo, Guangyu and Huang, Dong and Han, Junwei},
  booktitle={2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6762--6770},
  year={2018},
  organization={IEEE}
}