The Impact of Real Rain in a Vision Task

Abstract

Single image deraining has made impressive progress in recent years. However, the proposed methods are heavily based on high-quality synthetic data for supervised learning which are not representative of practical applications with low-quality real-world images. In a real setting, the rainy images portray a scene with a complex degradation caused by the rain weather and the low-quality factors. The goal of this paper is to investigate the impact of two visual factors that affect vision tasks: image quality and rain effect . To evaluate this, an image dataset with images varying these factors has been created. Aiming to evaluate them, different object detection algorithms are applied and evaluated on the dataset. Our findings indicate that the fine-tuned models can efficiently cope with this problem regardless of the rain intensity of the scene, however it is greatly affected by the image quality gap.