This repo includes the source code of the paper: "LSTM Pose Machines" (CVPR'18) by Yue Luo, Jimmy Ren, Zhouxia Wang, Wenxiu Sun, Jinshan Pan, Jianbo Liu, Jiahao Pang, Liang Lin.

Contact: Yue Luo ([email protected])

You can click the images below to watch our results on video-based pose estimation. The first one is the comparison with the state-of-the-art single image pose estimation method "Convolutional Pose Machines(CPMs)" on videos. Second one is our LSTM Pose Machines on video pose estimation.

The code is tested on 64 bit Linux (Ubuntu 14.04 LTS). You should also install Matlab (R2015a) and OpenCV (At least 2.4.8). We have tested our code on GTX TitanX with CUDA8.0+cuDNNv5. Please install all these prerequisites before running our code.

1. Get the code.

   git clone https://github.com/lawy623/LSTM_Pose_Machines.git
   cd LSTM_Pose_Machines

2. Build the code. Please follow Caffe instruction to install all necessary packages and build it.

   cd caffe/
   # Modify Makefile.config according to your Caffe installation/. Remember to allow CUDA and CUDNN.
   make -j8
   make matcaffe

3. Prepare data. We write all data and labels into .mat files.

• Please go to directory dataset/, and run get_data.sh to download PENN and JHMDB datasets.
• To create the .mat files, please go to directory dataset/PENN and dataset/JHMDB, and run the matlab scripts JHMDB_PreData.m and PENN_PreData.m respectively. It will take some time to prepare data.

• As described in our paper, we first trained a "single image model" based on the repository: Convolutional Pose Machines(CPMs). You can download this model at [Google Drive|Baidu Pan]. Put it in training/prototxt/preModel after downloading it. If you hope to train it by yourself, we also provide the prototxts in training/prototxt/preModel. You can train this model with our prototxts using the codes released by CPMs. This single image model is trained on LEEDS Sport Dataset and MPII Dataset.
• To train our LSTM Pose Machines on video datasets, go to training/ to run video_train_JHMDB.m or video_train_PENN.m. You can also run the matlab scripts from terminal at directorytraining/ by following commands. By default matlab is installed under /usr/local/MATLAB/R2015a. If the location of your matlab is not the same, please modify train_LSTM.sh if want to run the scripts from terminal. Notice that, if you want to train our LSTM Pose Machines on sub-JHMDB datasets, please modify line 10 of video_train_JHMDB.m and set the correct subset ID before your run this script.

   ## To run the training matlab scripts from terminal
   sh prototxt/PENN/LSTM_5/train_LSTM.sh   #To trained on PENN dataset
   ## Or
   sh prototxt/sub-JHMDB/LSTM_5_Sub1/train_LSTM.sh   #To trained on sub-JHMDB subset 1, change `line 10` of `video_train_JHMDB.m` to be `modelID = 1` first.
   sh prototxt/sub-JHMDB/LSTM_5_Sub2/train_LSTM.sh   #To trained on sub-JHMDB subset 2, change `line 10` of `video_train_JHMDB.m` to be `modelID = 2` first.
   sh prototxt/sub-JHMDB/LSTM_5_Sub3/train_LSTM.sh   #To trained on sub-JHMDB subset 3, change `line 10` of `video_train_JHMDB.m` to be `modelID = 3` first.

• Download our trained models from [Google Drive|Baidu Pan]. Put these models in model/PENN/ and model/sub-JHMDB/ respectively.
• Go to directory testing/. Specify the model ID you want to test by modifying line 15 of benchmark.m and setting the correct benchmark_modelID. Then you can run test_LSTM.sh which runs the matlab test script to get our evaluation results. Please look in test_LSTM.sh and modify the matlab bin location and -logfile name before running this script.
• Predicted results will be saved in testing/predicts/. You can play with the results by ploting predicted locations on images.
• Orders of the predicted accuracy for two datasets will be as follows:

  ## PENN Dataset
  Head    R_Shoulder L_Shoulder R_Elbow   L_Elbow   R_Wrist   L_Wrist   R_Hip    L_Hip    R_Knee    L_Knee   R_Ankle   L_Ankle || Overall
  98.90%  98.50%     98.60%     96.60%    96.60%    96.60%    96.50%    98.20%   98.20%   97.90%    98.50%   97.30%    97.70%  || 97.73%
  
  ## sub-JHMDB Dataset
  Neck     Belly    Head     R_Shoulder L_Shoulder  R_Hip    L_Hip    R_Elbow  L_Elbow  R_Knee   L_Knee  R_Wrist  L_Wrist  R_Ankle  L_Ankle || Overall
  99.20%   98.97%   98.27%   96.67%     96.13%      98.83%   98.63%   90.17%   89.10%   96.40%   94.80%  85.93%   86.17%   91.90%   89.90%  || 94.09%

To get the results in our paper, you need to remove unlisted joints, calculate average and reorder the accuracy.

• We provide the sample visualization code in testing/visualization/, run visualization.m to visually get our predicted result on PENN dataset. Make sure your have already run the testing script for PENN before visualizing the results.

Please cite our paper if you find it useful for your work:

@article{Luo2017LSTMPose,
    title={LSTM Pose Machines},
    author={Yue Luo, Jimmy Ren, Zhouxia Wang, Wenxiu Sun, Jinshan Pan, Jianbo Liu, Jiahao Pang, Liang Lin},
    journal={arXiv preprint arXiv: 1712.06316},
    year={2017},
}