HFGR: Hierarchical Fine‐Grained Regression for Visibility Estimation

ABSTRACT Accurate visibility estimation is crucial in enormous real‐world applications like transportation and autonomous driving. Recent image‐based solution has attracted more interests due to the low cost and promising performance. However, existing methods typically regards the visibility estimation as a simple regression problem, which often fails to capture the subtle variations of different visibility levels. To tackle the problem, in this paper, we propose a novel hierarchical fine‐grained regression (HFGR) model, which employs a multi‐level structure that categorizes visibility into coarse‐to‐fine ranges and refines these estimated values with a fine‐grained regression. Specifically, our designed hierarchical model perceives a global continuity weights, swiftly narrowing down the prediction range, while the small interval regression can also capture local features, achieving a satisfying fine‐grained perception. By the proposed hierarchical regression, we can differentiate the subtle varieties between different visibility levels. The proposed HFGR model achieves the significant performance improvements in different fog conditions on our collected real‐world highway foggy image dataset, offering a reliable solution for real‐world applications.