NTIRE 2022 Image Inpainting Challenge: Report

Andrés Romero, Angela Castillo, Jose Abril-Nova, Radu Timofte, Ritwik Das, Sanchit Hira, Zhihong Pan, Min Zhang, Baopu Li, Dongliang He, Tianwei Lin, Fu Li, Chengyue Wu, Xianming Liu, Xinying Wang, Yi Yu, Jie Yang, Rengang Li, Yaqian Zhao, Zhenhua Guo, Baoyu Fan, Xiaochuan Li, Runze Zhang, Zeyu Lu, Junqin Huang, Gang Wu, Junjun Jiang, Jiayin Cai, Changlin Li, Xin Tao, Yu-Wing Tai, Xiaoqiang Zhou, Huaibo Huang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 1150-1182

Image Inpainting has recently become an important research problem due to the rise of generative image synthesis models. While many solutions have been proposed for this problem, it is challenging to establish a testbed due to the different possible types of inpainting masks, e.g., completion mask, expand mask, thick brushes mask, etc. Most inpainting solutions shine on object removal or texture synthesis, while semantic generation is still difficult to achieve. To address these issues, we introduce the first general Image Inpainting Challenge. The target is to develop solutions that can achieve a robust performance across different and challenging masks while generating compelling semantic images. The proposed challenge consists of two tracks: unsupervised image inpainting and semantically-guided image inpainting. For Track 1, the participants were provided with four datasets: FFHQ, Places, ImageNet, and WikiArt, and trained their models to perform a mask-agnostic image inpainting solution. For Track 2, FFHQ and Places only. This report gathers the description and discussion of all solutions that participated in the final stage of the challenge.