Deep Pose

I think more tuning is possible. If you have some advice, please tell me!

Features

• Chainer implementation
• Subsequent stages
• Image viewer on web browsers. (Flask and flask-socketio are needed)

Supported Datasets

• [x] FLIC
• [ ] LPS

Testing Environments

Ubuntu 14.04

• Python 2.7
• Chainer 1.9.1
• OpenCV 2.4.8
• Flask 0.11.1
• Flask_SocketIO 2.4

Arch Linux

• Python 3.5
• Chainer 1.9.1
• OpenCV 3.1.0
• Flask 0.10.1
• Flask_SocketIO 2.2

Windows 7 (not recommend)

• Python 2.7
• Chainer 1.9.1
• OpenCV 3.1.0
• Flask 0.11.1
• Flask_SocketIO 2.5

Training

First, download FLIC FULL and FLIC PLUS to some directory, and set the path to settings.json. And also fix CASCADE_PATHS to detect faces for your environment.

To start training, please execute the following command.

./scripts/train.py --stage 0

For subsequent stage training, --joint_idx argument is needed.

./scripts/train.py --stage 1 --joint_idx 0
./scripts/train.py --stage 1 --joint_idx 1
./scripts/train.py --stage 2 --joint_idx 0  # and so on

--resume argument is also supported.

To check current training state, please open http://localhost:8889/ (port number can be changed by settings.json). The error rate graph and visualized images can be seen.

If you want to use GPU, please set GPU parameter in settings.json to a positive number.

Use trained models

Execute the following command, and open http://localhost:8889/.

./scripts/use_model.py

Settings is common with training (settings.json).

Results

First stage

<img src="https://raw.githubusercontent.com/takiyu/DeepPose/master/screenshots/first_stage_result.jpg">

Subsequent stages are training now.

Async Mode Setting

This project uses Python threading or multiprocessing package and it can be configured by ASYNC_MODE in settings.json. On Linux process mode is better due to the speed, but on Windows only thread mode is valid.

TODO

• [ ] Tune training parameters (learning rate, bounding box sigma and so on).
• [ ] Replace multiprocess.Queue and Event to threading's ones on the thread mode.