Josephine Sullivan: Football Player 3D Pose Reconstruction from Visual Data

In this talk I will present our current work and results at CVAP on the estimation of a human's 3D pose - football players during matches - from visual images. It is a challenging problem due to the high degree of articulation of a human skeleton and the large number of self-occlusions that occur when a person is imaged. However, we exploit multi-view footage to help resolve these ambiguities.

Our most recent research builds on existing algorithms and systems in the literature - the “Microsoft Kinect system” and “pictorial structures” - and tries to marry their advantages. The former solves the skeleton estimation problem in home environments using decision trees, applied to depth camera measurements, learnt from abundant training data. While the model based “pictorial structures” only requires limited training sets and is flexible enough to recognize poses outside the training set, but comes at a cost of frequently hallucinating invalid poses.

These two approaches highlight the current division within computer vision of whether one should build discriminative or generative systems to enable recognition and estimation.