Markerless Motion Capture in the Crowd

TL;DR

This paper presents a crowdsourcing approach to obtain 2D motion capture data from videos, optimizing accuracy and efficiency, and reconstructing 3D motion structures from the collected data.

Contribution

It introduces a novel crowdsourcing method for motion capture that combines human input with reconstruction techniques to derive 3D motion from video recordings.

Findings

Effective crowdsourcing strategies for motion tracking.

High-quality 2D motion data obtained from crowd workers.

Successful reconstruction of 3D motion structures.

Abstract

This work uses crowdsourcing to obtain motion capture data from video recordings. The data is obtained by information workers who click repeatedly to indicate body configurations in the frames of a video, resulting in a model of 2D structure over time. We discuss techniques to optimize the tracking task and strategies for maximizing accuracy and efficiency. We show visualizations of a variety of motions captured with our pipeline then apply reconstruction techniques to derive 3D structure.