RepFinder: Finding Approximately Repeated Scene Elements for Image Editing

摘要

Repeated elements are ubiquitous and abundant in both manmadernand natural scenes. Editing such images while preserving the repetitionsrnand their relations is nontrivial due to overlap, missing parts,rndeformation across instances, illumination variation, etc. Manuallyrnenforcing such relations is laborious and error-prone. We proposerna novel framework where user scribbles are used to guide detectionrnand extraction of such repeated elements. Our detection process,rnwhich is based on a novel boundary band method, robustly extractsrnthe repetitions along with their deformations. The algorithm onlyrnconsiders the shape of the elements, and ignores similarity basedrnon color, texture, etc. We then use topological sorting to establishrna partial depth ordering of overlapping repeated instances. Missingrnparts on occluded instances are completed using information fromrnother instances. The extracted repeated instances can then be seamlesslyrnedited and manipulated for a variety of high level tasks thatrnare otherwise difficult to perform. We demonstrate the versatility ofrnour framework on a large set of inputs of varying complexity, showingrnapplications to image rearrangement, edit transfer, deformationrnpropagation, and instance replacement.