Estimating visibility via differential regression network

Estimating visibility via differential regression network

연구 분야: Networking

학회: Multimedia Systems

The high risk of traffic accidents on highways caused by low visibility makes it vital for accurate visibility estimation. As an effective and low-cost solution, image-based visibility estimation has achieved great progress. However, most existing methods simply regress the visibility from a single image, while their predictions suffer from significant differences for some images with similar visibility levels due to large variations such as diverse scenes, changing lighting and seasons. On the other hand, these methods may also produce similar visibility values for the images captured by the same cameras, despite varying visibility conditions. To address the above issues, we argue that the key is to discover the differences among the images and predict visibility based on the differences. Inspired by this, in this paper, we propose a novel end-to-end differential regression network designed to estimate the visibility differences between pairs of similar images, rather than estimating each image separately. Our proposed method allows the model to concentrate on visibility-related features by capturing the discrepancy between image pairs, thereby minimizing the impact of large image variations. For training and evaluation, we construct two comprehensive and realistic datasets, JS-FHVI and DG-FHVI, collected from real highway surveillance videos. The comprehensive experiments show the effectiveness and superiority of our proposed method.