silasxue / deeppose-1

DeepPose implementation with TensorFlow

DeepPose

NOTE: This is not official implementation. Original paper is DeepPose: Human Pose Estimation via Deep Neural Networks. SECOND NOTE: This implementation was a project for my Pattern Recognition Course at METU. Codes are in a very primitive level. But people might find them useful.

Requirements

• TensorFlow (Google's Neural Network Toolbox)
• Python 2.7.x
• Numpy
• SciPy.io (For loading .mat files)
• PIL or Pillow
• Six (Python 2 and 3 Compatibility Library)

Important

Edit values in 'LSPGlobals.py' as you want them. All the codes run using values in that file.

Data preparation

python GetLSPData.py

This script downloads Leeds Sports Pose Dataset (http://www.comp.leeds.ac.uk/mat4saj/lsp.html) and performs resizing as your Neural Network input size. Resized images and their labels are saved into binary files.

Dataset:

• LSP Dataset

Start training

Just run:

python TrainLSP.py

To Follow Progress

tensorboard --logdir=/path/to/log-directory   #path is '~/Desktop/LSP_data/TrainData' if LSPGlobals.py is unchanged

Evaluating the trained model

python EvalDeepPose.py

This will get all images placed in '--input_dir' with extension '--input_type' will draw stick figures on images based on estimations from the model. Drawn images will be placed in '--output_dir'.

Video

I recommend you to use ffmpeg to turn videos into images, feed them to network and make video from the drawn images using ffmpeg again.

Here is my example