Alternating exposure control network for real-world environments

Real-world environments exhibit a large dynamic range that far exceeds the capabilities of conventional cameras in a single capture. Particularly in autonomous driving scenarios, improper exposure settings degrade the image quality and the performance of downstream high-level tasks, which pose significant challenges to existing exposure control methods. However, when a captured image contains textureless and reflective regions, existing methods produce compromised global exposure estimates, which result in suboptimal predictions for critical objects. When encountering illumination changes, existing methods demonstrate delayed adjustments that only occur after the image is saturated, which leads to consecutive saturated frames. To address these issues, we propose a novel alternating exposure control network to achieve a larger dynamic range via alternating exposure settings across adjacent frames. Specifically, to achieve accurate exposure prediction in critical regions, we leverage semantic-aware weight maps to explicitly supervise the network for producing pixel-wise exposure values. The weight map highlights local regions with rich semantic information and assigns higher weights to the corresponding exposure values. To maintain sensitivity for illumination variation conditions, we introduce hierarchical clustering to obtain global illumination features. By aggregating features across different illumination levels, we construct hierarchical representations to produce global exposure shifts. Our exposure control network is jointly trained with an object detector and implemented for lightweight deployment. Extensive validation in real-world illumination variation scenarios demonstrates the ability to enhance both image quality and downstream task performance.