August Bergöö och Oliver Lundgren presenterar sitt examensarbete "Smoothing techniques for 3D animated human pose estimation data"

Titel: Smoothing techniques for 3D animated human pose estimation data

Abstract:

Virtual 3D reconstructions of live sport events are on the horizon and to produce a high quality experience for viewers it is important that the movements of the 3D models look natural. Today, state of the art pose estimators produces data that con- tains noisy data, resulting in jittery animations with pose errors. The goals for this thesis is to evaluate the performance on an LSTM and classical filters in their ability to reduce such noise, and if they can be used to improve the viewer experience. This was done by comparing the Mean Per Joint Positional Error (MPJPE) and Absolute Accerational Error (AAE) for artificially added noise to motion capture data and comparing the results. A qualitative survey was also constructed to evaluate how the different methods affected the human perception of the animations when focus- ing on different aspects. We concluded that our methods were (to varying degrees) able to reduce the amount of noise introduced to the motion data. The qualitative study also indicated that the Savitzky-Golay smoothing algorithm improved the par- ticipants perception of the animations.

Examinator: Kalle Åström

Handledare: Håkan Ardö och Magnus Oskarsson