Retargeting HR Aerial Photos Under Contaminated Labels With Application in Smart Navigation
Luming Zhang,Ming Chen,Bing Tu,Yi Li,Yingjie Xia
DOI: https://doi.org/10.1109/tits.2023.3288877
IF: 8.5
2023-01-01
IEEE Transactions on Intelligent Transportation Systems
Abstract:Retargeting aims to shrink a photo wherein the perceptually prominent regions are appropriately kept. In practice, optimally shrinking a high resolution (HR) aerial photo is a useful tool for smart navigation. Nowadays, vehicle drivers’ path planning is generally guided by an HR aerial photo recommended by a navigation App like Google Maps. Owing to the limited and various resolution of vehicle displays, we have to retarget each original HR aerial photo accordingly, wherein the navigation-aware regions can be well preserved. In practice, HR aerial photo retargeting is non-trivial due to three challenges: 1) the rich number of internal objects and their complex spatial layouts, 2) deriving the region-level semantics from potentially contaminated image labels, and 3) the inefficiency of retargeting each HR aerial photo with millions of pixels. To handle these problems, we propose a novel HR aerial photo retargeting pipeline that can intelligently avoid the negative effects from incorrect image labels. The key is a noise-tolerant hashing algorithm that converts image-level semantics into the hash codes corresponding to different regions, which guides the HR aerial photo shrinking. More specifically, for each HR aerial photo, we extract visually/semantically salient object patches inside it. To explicitly encode their spatial layout, we construct a graphlet by linking the spatially adjacent object patches into a small graph. Subsequently, a binary matrix factorization (MF) is designed to exploit the underlying semantics of these graphlets, wherein three attributes: i) binary hash codes learning, ii) noisy labels refinement, iii) deep image-level semantics, are collaboratively encoded. Such binary MF can be solved iteratively and each graphlet is subsequently converted into the binary hash codes. Finally, the hash codes corresponding to graphlets within each HR aerial photo are utilized to learn a Gaussian mixture model (GMM) that optimizes the HR aerial photo retargeting. During the experimental validation, we compiled a smart navigation dataset including 132743 planned paths annotated from 10132 HR aerial photos, based on which comparative study has demonstrated the superiority of our method.
engineering, electrical & electronic,transportation science & technology, civil