idotc/DLMSCM

Human Pose Estimation with Deeply Learned Multi-Scale Compositional Models

This implementation is based on the code and data from [1-7]. We thank all the authors for kindly sharing these valuable resources.

Setting

1. Install Torch

• Option (a): Follow http://torch.ch/docs/getting-started.html
• Option (b): Use the Docker image kaixhin/cuda-torch:8.0 (https://hub.docker.com/r/kaixhin/cuda-torch/)

2. Install dependences

apt-get install libhdf5-serial-dev
luarocks install hdf5

3. Prepare datasets Download MPII [3] and LSP [4] datasets and create symbolic links so that their respective JPEG images can be found in:

data/mpii/images
data/lsp_dataset/images
data/lspet_dataset/images

Training

Train a model with 3 semantic levels on 1 GPUs(16G).

./experiments/PLACEHOLDER/train_dlmscm_l3.sh

where PLACEHOLDER can be:

• mpii: Train with MPII training data excluding the 3K validation samples.
• mpii-include-val: Train with MPII training data including the 3K validation samples.
• mpii-lsp: Train with MPII training data and corrected LSP training data.

Testing

1. Download trained models from our project website and put them in checkpoints/saved.

2. Get human pose predictions using a trained model with 3 semantic levels

./experiments/PLACEHOLDER/predict_dlmscm_l3.sh

where PLACEHOLDER can be:

• mpii: Predict on MPII 3K validation samples.
• mpii-include-val: Predict on MPII testing data.
• mpii-lsp: Predict on LSP testing data.

3. Evaluate the predictions by comparing them against the corresponding ground truth.

• Check http://human-pose.mpi-inf.mpg.de/#evaluation for evaluation on MPII data.
• You may evaluate the PCK@0.2 scores of your model on the LSP test set. To get start, download our prediction pred_multiscale_1_best.h5 and eval code from github, and run the MATLAB script transfer_pre_test.m.

References

[1] Wei Tang, Pei Yu, and Ying Wu. "Deeply Learned Compositional Models for Human Pose Estimation." in Proceedings of European Conference on Computer Vision (ECCV'18), Munich, Germany, Sept. 2018.

[2] Wei Yang, Shuang Li, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. "Learning feature pyramids for human pose estimation." In ICCV 2017.

[3] Alejandro Newell, Kaiyu Yang, and Jia Deng. "Stacked hourglass networks for human pose estimation." In ECCV 2016.

[4] https://github.com/facebook/fb.resnet.torch

[5] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. "2d human pose estimation: New benchmark and state of the art analysis." In CVPR 2014.

[6] Sam Johnson and Mark Everingham. "Clustered pose and nonlinear appearance models for human pose estimation." In BMVC 2010.

[7] Ben Sapp and Ben Taskar. "Modec: Multimodal decomposable models for human pose estimation." In CVPR 2013.