Human Pose Estimation with Iterative Error Feedback (IEF)

This package is an implementation of the algorithm described in our paper (pdf) for estimating human pose in natural images.

Requirements

• Python 2.7
• OpenCV python interface (python-opencv)
• Caffe

Currently, it is necessary to use caffe version at the link. A patch that you can apply to your own caffe repository will be provided shortly in the future.

Installation

This package has been tested on Ubuntu 14.04.

• Download IEF models and supporting code: ./setup.sh

Running Demo

In the main code directory launch ipython and run,

from src import test_demo as td
#Define pose-predictor class
ief    = td.PoseIEF()
#Name of the image
imName = 'src/test_images/elvis.jpg'
#Point (x,y) on the torso of the person whose pose is to be estimated
bodyPt = (108, 98)
#Predict the pose
pose,_ =  ief.predict(imName, bodyPt)
#Visualize the result
import scipy.misc as scm
im = scm.imread(imName)
td.vis.plot_pose_stickmodel(im, pose.squeeze().transpose((1,0)))

Note: This code only runs 1 image in a single batch and is hence runs slower than what can be achieved with larger batch sizes.

Downloading MPII Annotations in python

See the wiki page.

Coming Soon

The README will be shortly updated with more details.