Title:

Camera Trajectory Modeling from Soccer Video

Abstract:

How do professional camera operators move the camera when they film a sports game? Can we capture these natural dynamics, and learn to control a Pan-Tilt-Zoom-camera in a human-like way? We approach this with a pipeline that extracts a camera parameter time-series directly from recorded soccer videos, and a discrete autoregressive modeling framework. We compare some discretization schemes and apply a classic model from Natural Language Processing, which in the end can sample a diversity of camera motions. We also take a look at reference based metrics, measuring the "realism" of generated camera motion trajectories.

Speaker: Markus Salzer