Illuminator:Image-based illumination editing for indoor scene harmonization  

作　　者：Zhongyun Bao Gang Fu Zipei Chen Chunxia Xiao 

机构地区：[1]School of Computer Science,Wuhan University,Wuhan 430072,China

出　　处：《Computational Visual Media》2024年第6期1137-1155,共19页计算可视媒体(英文版)

基　　金：supported by the National Natural Science Foundation of China(61972298 and 62372336),CAAI–Huawei MindSpore Open Fund,and Wuhan University–Huawei GeoInformatices Innovation Lab.

摘　　要：Illumination harmonization is an important but challenging task that aims to achieve illumination compatibility between the foreground and background under different illumination conditions.Most current studies mainly focus on achieving seamless integration between the appearance(illumination or visual style)of the foreground object itself and the background scene or producing the foreground shadow.They rarely considered global illumination consistency(i.e.,the illumination and shadow of the foreground object).In our work,we introduce“Illuminator”,an image-based illumination editing technique.This method aims to achieve more realistic global illumination harmonization,ensuring consistent illumination and plausible shadows in complex indoor environments.The Illuminator contains a shadow residual generation branch and an object illumination transfer branch.The shadow residual generation branch introduces a novel attention-aware graph convolutional mechanism to achieve reasonable foreground shadow generation.The object illumination transfer branch primarily transfers background illumination to the foreground region.In addition,we construct a real-world indoor illumination harmonization dataset called RIH,which consists of various foreground objects and background scenes captured under diverse illumination conditions for training and evaluating our Illuminator.Our comprehensive experiments,conducted on the RIH dataset and a collection of real-world everyday life photos,validate the effectiveness of our method.

关 键 词：indoor scene illumination harmonization object illumination editing seamless integration shadow residual generation 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]