reduce resolution to make “small motion” assumption hold - since they move smaller pixel distance - allows optical flow and lucas-kanade with less errors upsampling = scaling u and v values from iterative lucas-kanade
