Engineering Commons^™

Ghost Towns: Semantically Labelled Object Removal From Video, William Clifford, Charles Markham

Session 4: 2D, 3D Scene Analysis and Visualisation

This paper describes a method used to produce a video of a road in which the foreground itemswhich obstruct the view of the road have been removed i.e. other vehicles. Once these regions have been identified they are replaced using suitable images that closely resemble the original background. The work considers an approach that uses multiple video sequences of the same road (C1...Cn). One video is identified as video Cp , that requires the least repair. All instances of vehicles in each frame of video were identified using a Convolutional Neural Network (CNN). The regions associated with each vehicle were …