Image Inpainting in Generative Adversarial Nets Combining with Depth Information

Zidong Lin


Supervised by Yukun Lai; Moderated by Yipeng Qin

Digital images are versatile nowadays. However, captured images may not contain the exact content of interest. For example, an image may contain extra objects/persons that are not intended to be included. It is therefore highly demanding to develop more intelligence image manipulation techniques. Recent advances in deep learning and generative adversarial networks provide a much more powerful tool for image manipulation.

Recently, with the development of monocular depth estimation, it is possible to extract the depth information from single RGB images and apply it to image inpainting.This project investigates how to combine the depth information of images with image inpainting and assess the results quality.

Final Report (20/09/2022) [Zip Archive]

Publication Form