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:


How to start

Download models

Put downloaded models in the models/weights folder.

Run demo scripts for body pose estimation and for hand pose estimation. I take 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.

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

Build your own model

Create dataset

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

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/ for example.

Please see 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 for more details.


This project is licensed under the Apache 2.0 License.