Supervisors: Dr. Danda Pani Paudel, Dr. Zhiwu Huang, and Prof. Luc Van Gool
With the advent of GAN, significant progress has been made in various image manipulation tasks, including image inpainting. However, the image objects removal, a widely used manipulation task, has received very little attention; mainly due to the difficulty of obtaining paired or unpaired examples. In this work, we proposed a method that realistically removes objects of various sizes from a high resolution image, without using any example-based supervision. Our method exploits the GAN to generate high quality images, which is the same as the input but the targeted objects removed. We treat the object removal as a task of conditional inpainting where the inpainted region must not reproduce the targeted object or any of its kind. We achieve this with two key contributions: (a) utilizing a pixel-wise GAN for consistent inpainting; and (b) proposing a patch-based GAN that exploits the neighborhood context to reconstruct realistic images, using only a weak supervision. We experimentally show that the proposed framework can successfully remove objects of various sizes in a high-resolution image, using only the concept of objects’ semantics.