3D Reconstructions Applied on Broadcast Sports

Alexandru Ionascu, West University of Timisoara

Abstract: All the major broadcast partners provide 360 views in all the leading sports and further interactive scene synthesis for commentary and analysis. In addition, many triple-A video game studios use markerless optical tracking to create new animations. In this talk, we will explore the possibilities by looking at broadcast-monocular football and tennis videos without additional sensors or data. To discuss the current state-of-the-art method, we typically refer to (1) complete 3D scenes with human body regression and neural rigging from video or (2) pseudo-3D environments with neural rendering pipelines (pix2pix or vid2vid). We will also introduce a sampling-based approach for video processing, and we can experiment with this technique for athlete 3D reconstruction on existing and new poses.