RORD: A Real-world Object Removal Dataset
Min-Cheol Sagong (Korea Univ.),* Yoon-Jae Yeo (Korea Univiersity), Seung-Won Jung (Korea University), Sung-Jea Ko (Korea University)
The 33rd British Machine Vision Conference
Abstract
Various convolutional neural networks (CNNs)-based image inpainting techniques have been actively studied to remove unwanted objects or restore missing parts in recent years. The common standard for training image inpainting CNNs is synthesising hole regions on the existing datasets, such as ImageNet and Places2. However, from the viewpoint of the object removal task, such a methodology is suboptimal because actual pixels behind objects, i.e., ``ground truth'', cannot be used for training. Facing this problem, we introduce Real-world Object Removal Dataset (RORD), a large-scale collection of image pairs with and without objects. RORD consists of a wide range of real-world scenes, plus two types of pixel-accurate annotations, i.e., object mask and segmentation map. Our dataset allows existing image inpainting models to be trained accurately as well as evaluated with high confidence. In this paper, we describe in detail how the dataset is constructed and demonstrate the validity and usability of RORD. RORD is publicly available at (Owing to the blind review policy, we omit URL. It will be available upon acceptance of the manuscript.)Video
Citation
@inproceedings{Sagong_2022_BMVC,
author = {Min-Cheol Sagong and Yoon-Jae Yeo and Seung-Won Jung and Sung-Jea Ko},
title = {RORD: A Real-world Object Removal Dataset},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0542.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by