Bilkent University
Department of Computer Engineering

---

CS 590/690 SEMINAR

 

Conditional Image Inpainting with Style Codes

 

Hakan Sivük

Master Student
(Supervisor: Asst.Prof.Ayşegül Dündar Boral )
Computer Engineering Department
Bilkent University

Abstract: Image inpainting is studied in many research projects with various techniques. It basically aims to fill in missing parts of an image in a realistic way. In conditional image inpainting, there is a precondition that is used while filling in the missing pixels. In this project, I aimed to use instance maps as a precondition to generate missing parts consistent with object boundaries. Style codes are encoded from valid parts of the image and perform instance-wise pooling to complete the image with a consistent style. Additionally, an algorithm is proposed where we can use the style code of the same object encoded from another image for generating masked areas. It prevents using style codes with no object information encoded in them, especially for areas with fully masked objects. Besides, it provides diversity where we can fill masked objects with different styles.

 

DATE: April 3, Monday @ 16:50 Place: Zoom