A Lightweight Framework for Objection Detection in Adverse Lighting Conditions

Although the latest object detection methods have demonstrated strong performance on large-scale comprehensive datasets, various adverse lighting conditions limit their potential application in real-world scenarios, especially the deployment on mobile robots. In this study, we propose a novel and lightweight re-illumination framework for end-to-end adaptive enhancement for adverse lighting images. Specifically, we employ a set of differentiable image processing modules and pixel-wise curve parameter mapping to adapt to various lighting conditions. We use YOLOv3 detection loss to learn the curve parameters of the U-shaped parameter predictor (UPP) in a weakly-supervision manner. We further deploy the proposed framework to a low-power hardware platform. The experimental results demonstrate the effectiveness of our proposed method in various adverse lighting conditions(i.e. haze, low-light).