RGB-D Co-Segmentation on Indoor Scene with Geometric Prior and Hypothesis Filtering
Indoor scene parsing is crucial for applications like home surveillance systems.Although deep learning based models like FCNs 10 have achieved outstanding performance,they rely on huge amounts of hand-labeled training samples at pixel level,which are hard to obtain.To alleviate labeling burden and provide meaningful clues for indoor applications,its promising to use unsupervised co-segmentation methods to segment out main furniture,such as bed and sofa.Following traditional bottom-up co-segmentation framework for RGB images,we focus on the task of co-segmenting main furniture of indoor scene and fully utilize the complementary information of RGB-D images.First,a simple but effective geometric prior is introduced,using bounding planes of indoor scene to better distinguish between foreground and background.A two-stage hypothesis filtering strategy is further integrated to refine both global and local object candidate generation.To evaluate our method,the NYUD-COSEG dataset is constructed,on which our method shows significantly higher accuracy compared with previous ones.We also prove and analyze the effectiveness of both bounding plane prior and hypothesis filtering strategy with extensive experiments.
Indoor RGB-D co-segmentation Geometric prior for indoor scene Object hypothesis generation
Lingxiao Hang Zhiguo Cao Yang Xiao Hao Lu
National Key Lab of Science and Technology of Multispectral Information Processing,School of Automation,Huazhong University of Science and Technology,Wuhan 430074,Hubei,China
国际会议
广州
英文
168-179
2018-11-23（万方平台首次上网日期，不代表论文的发表时间）