/ Machine Learning

Tensorflow implementation of Convolutional Pose Machines

Tensorflow implementation of Convolutional Pose Machines

Convolutional Pose Machines - Tensorflow


This is the Tensorflow implementation of Convolutional Pose Machines, one of the state-of-the-art models for 2D body and hand pose estimation.

Tracking support for single hand.

With some additional features:

  • Easy multi-stage graph construction
  • Kalman filters for smooth pose estimation
  • Simple self-tracking module


  • Windows 10 / Ubuntu 16.04
  • Tensorflow 1.4.0
  • OpenCV 3.2

How to start

Download models

Put downloaded models in the models/weights folder.

Run demo scripts

demo_cpm_body.py for body pose estimation and demo_cpm_hand.py for hand pose estimation.
I take demo_cpm_hand.py for example.

First set the DEMO_TYPE. If you want to pass an image, then put the path to image here.
If you want a live demo through a webcam, there are few options.

  • MULTI will show multiple stages output heatmaps and the final pose estimation simultaneously.
  • SINGLE will only show the final pose estimation.
  • HM will show each joint heatmap of last stage separately.

You can also use video files like .avi, .mp4, .flv.

The CPM structure assumes the body or hand you want to estimate is located in the middle of the frame.
If you want to avoid that, one way is to add a detector at the begining, and feed the detected bounding box image into this model.

Run hand demo with tracker

Download tf-version hand pose model and execute run_demo_hand_with_tracker.py

Build your own model

Create dataset

See utils/create_cpm_tfr_fulljoints.py for an example.
If you want to follow the script, you need to prepare your data like

  • dataset/person_0/imgs/
  • dataset/person_0/labels.txt

And in labels.txt, the data format is
imgs_0.jpg bbox_top_left_y bbox_top_left_x bbox_bot_right_y bbox_bot_right_x joint_0_y joint_0_x joint_1_y joint_1_x ....

Original Training

See models/nets for model definition, I take models/nets/cpm_hand.py for example.

  • Create a model instance
  • Set how many stages you want the model to have (at least 2)
  • Call build_loss if you want to do the training
  • Use self.train_op to optimize the model

Please see run_training.py for an example.

Distillation Training

I implement a distillation training scheme, which is a teacher-student learning process.
It suppose you already have a strong teacher network, and want to train a smaller student network that can perform like its teacher.
See run_training_distillation.py for more details.