Bottom-Up and Top-Down Reasoning with Hierarchical Rectified Gaussians

Instructions:

• Compile MatConvNet (If you use CuDNN, make sure you set the $PATH and $LD_LIBRARY_PATH correct)
• Download the pre-trained vgg-16 model from MatConvNet and put in matconvnet/
• Download MPII and put in data/mpii_human/
  • see data/mpii_human/README for more details
• Call run_qp1_mpii.m for training & testing a qp1 model
• Call run_qp2_mpii.m for training & testing a qp2 model

Pretrained models:

Download our trained models here.

Notes:

A few necessary changes are made based on the original MatConvNet.

See our code and models for face landmark localization on AFLW dataset.

Credits:

The implementation is based on & inspired by MatConvNet and MatConvNet-FCN.

Thanks James for dag_viz.m which dumps a MatConvNet model into a dot file.