MoDeep: A Deep Learning Framework for Human Pose Estimation

Illustration: NYU

Accurate identification of people’s pose in video is of great importance. Its applications include gesture-based controls such as Kinect and motion capture systems without markers. New York University researchers recently developed a deep learning architecture using both color and motion features for human pose estimation.

The deep learning framework, named MoDeep, based on a multi-resolution convolutional network, was published in a recent paper "MoDeep: A Deep Learning Framework Using Motion Features for Human Pose Estimation". The study has also proposed new motion features and created a new dataset called FLIC-motion by augmenting the Frames Labeled In Cinema (FLIC) dataset with the proposed motion features. According to the paper, MoDeep has been tested on the FLIC-motion dataset and outperforms existing state-of-the-art techniques for the task of human body pose detection in video.

For more details about MoDeep, please visit http://cs.nyu.edu/~ajain/accv2014/.