Superimposing perimeter banners dynamic content on soccer game videos pre-recorded from the main camera feed

Abstract

Advertisement billboards around a sports field have a fixed frame but most of the time a dynamic content. This makes a precise segmentation complex. This thesis comprises the developments of (1) a neural network capable to precisely segment billboards in a single image, (2) a temporally consistent camera calibration procedure able to reproject the billboard on the field, and (3) a compositing method for a seamless superimposition of the reprojected billboard on the field. Non-public annotations were used to train and test algorithms.