Scott Spurlock

Associate Professor, Computer Science, Elon University

Gamesourcing to Acquire Labeled Human Pose Estimation Data


Pose Estimation Examples

People

Ayman Hajja    Richard Souvenir    Scott Spurlock   

Overview

Human pose estimation from 2D images is a difficult problem. Most recent approaches are learning-based algorithms that require large data sets for training and testing. Creating these data sets can be expensive, slow, and tedious due to manual annotation. By using a gamesourcing approach, we can distribute the effort of annotating data among the players of a game, which makes the process both quicker and significantly more enjoyable.

Where previous gamesourcing research has focused on the creation of a specific game to solve a specific problem, in this work we propose a model that can work with any PC-based Microsoft Kinect game. Our framework runs side-by-side with any such game, capturing images and associated human joint locations to create a labeled data set appropriate for 2D pose estimation. Real-time filtering insures that only suitable samples are retained for the data set, filtering out instances of poor synchronization between RGB images and estimated joint location, inaccurate joint locations, and redundant poses which are highly similar to poses already captured.

We evaluated the method by using gamesourced data to train and test a recent 2D pose estimation algorithm. Results show that data collected in this manner is just as suitable as existing, manually annotated data sets for use as training data, while being much easier to procure.

Pipeline

Citation

Scott Spurlock and Richard Souvenir. "An evaluation of gamesourced data for human pose estimation," ACM Transactions on Intelligent Systems and Technology (TIST), 6(2), 2015.

Richard Souvenir, Ayman Hajja, Scott Spurlock. "Gamesourcing to Acquire Labeled Human Pose Estimation Data," Computer Vision and Pattern Recognition Workshops (CVPRW), 2012 IEEE Computer Society Conference on.