Recurrent 3D pose Sequence Machines.

By Mude Lin, Liang Lin, Xiaodan Liang, Keze Wang and Hui Cheng.

Introduction

Estimating 3D human pose from monocular images has many applications, includings human computer interaction, virtual reality, motion-sensing games and so on. We propose a novel Recurrent 3D Pose Sequence Machines(RPSM). You can use the code to train/evaluate a network for 3D pose estimation task. For more details, please refer to our paper.

Contents

1. Installation
2. Preparation
3. Train/Eval
4. Models

Installation

Pre-require

1. Ubuntu 14.04
2. NVIDIA GPU with 6GB graphic memory
3. Torch
4. OpenCV 2.4.9 pyhon interface
5. torch-hdf5
6. nngraph

Get the code. We will call the directory that you cloned into $RPSM_ROOT

https://github.com/Geekking/RPSM.git
cd RPSM

Preparation

Please see util/preprocess

Train/Eval

1. Train your model and evaluate the model.

1.1 Pretrain shared 2D pose module with MPII data.

The shared 2D pose module is trained with CPM code. and converted to torch module, we have provided a model which are stored at models/torch_model/caffe_d2_pose_module_shared.t7 in our Release models. You could unzip our provided model and run cp models/torch_model/caffe_d2_pose_module_shared.t7 $RPSM_ROOT/models/torch_model/caffe_d2_pose_module_shared.t7

1.2 main training phase

RPSM with 3 stages versions

``` Shell
cd $RPSM_ROOT/scrpts/rpsm and run bash train_rpsm_pretrained_rho3_t5.sh
After 50 epoches, the MPJPE metircs should be about 73. 
```

the snapshots will be stored in exp/h3m/rpsm_1024_rho3_t5

2. Evaluate snapshots.

``` Shell
Modify the `refineModel` parameter to you model in scripts/rpsm/test_rpsm_rho3.sh at line 17.
and `cd scripts/rpsm/  && bash test_rpsm_rho3.sh`

```

Models

Download trained model from Baidu Yun, and cd scripts/rpsm/ && bash test_rpsm_rho3.sh.

Our predicted result on Human 3.6M dataset could be found at Baidu Yun

Citation

If you like this work, please consider citing:

@inproceedings{linCVPR17RPSM,
    title = {Recurrent 3D Pose Sequence Machines},
    author = {Mude Lin and Liang Lin and Xiaodan Liang and Keze Wang and Hui Chen},
    booktitle = {CVPR},
    year = {2017}
}