RMPE: Regional Multi-person Pose Estimation(SPPE+SSTN part)

This repository includes Torch code for training our SPPE+SSTN model presented in our paper

Installation
Preparation
Evaluate
Train
Acknowledgements

Installation

To run this code, the following must be installed:

Torch7
hdf5 (and the torch-hdf5 package)
cudnn
qlua (for displaying results)
STN module(use this stable repo)

After all the installation is done, get the code. We will call the directory that you cloned this repo into $SPPE_ROOT

git clone https://github.com/Fang-Haoshu/multi-human-pose.git
cd multi-human-pose

Preparation

For evaluation only

Download pre-trained SPPE+SSTN model(Baidu cloud). By default, we assume the models are stored in $SPPE_ROOT/predict/.

For training

Download base networks and 4-stack parallel SPPE. By default, we assume the models are stored in $SPPE_ROOT/train/src.
Download MPII images. By default, we assume the images are stored in /data/MPII/images/.

Evaluate

You will need to generate bounding box first. Here we have already generated the bounding boxes in $SPPE_ROOT/predict/annot/mpii-test0.09/. To generate it yourself, please follow the guidlines in the main repo.

cd $SPPE_ROOT/predict
# make a soft link to the images
ln -s /data/MPII/images/ data/images
# get the predicted results
th main.lua predict-test

You can also use the following command to visualize the single person pose estimation results.

  qlua main.lua demo

Train

We finetune our model based on the pre-trained stacked-hourglass model.

cd $SPPE_ROOT/train/src
ln -s /data/MPII/images ../data/mpii-box/images
# First finetune the model using PGPG
th main.lua -expID finetune -usePGPG
th main.lua -expID finetune -usePGPG -continue -LR 0.5e-4 -nEpochs 10
# Then add parallel SPPE and SSTN on the finetuned model
# It should reach a final mAP of 80.*
th main.lua -expID final_model -loadModel '../exp/mpii-box/finetune/final_model.t7' -LR 0.5e-4 -addParallelSPPE -addSSTN -usePGPG
th main.lua -expID final_model -continue -nEpochs 10 -LR 0.1e-4 -addParallelSPPE -addSSTN -usePGPG

Acknowledgements

Thanks to Wei Liu, Alejandro Newell, Pfister, T., Kaichun Mo, Maxime Oquab for contributing their codes. Thanks to the authors of Caffe and Torch7!

Fang-Haoshu/multi-human-pose