Igor Susmelj

Master Thesis
Supervisors: Anna Volokitin, Eirikur Agustsson, Dr. Radu Timofte

Context-Aware Placement

Inserting new objects into images has long been the expertise of artists working with specialized editing software. This work consists of extracting an object from an existing image and then placing the extracted object into a new image. Recent works in computer vision have shown promising methods to solve the first problem using deep neural networks. The latter problem, however, remains an unsolved problem. In this thesis, we propose a novel system to automate this process, by both selecting a plausible location for an object to be placed in a scene, and also the best-suited size of the object. Our placement proposal network is based on using the context around objects as a detection feature. Results on the COCO dataset show that we are able to propose sensible locations and sizes better than baseline methods.