Domain Adaptation for Vehicle Detection Under Adverse Weather

Abstract

The images captured under varying illumination or adverse weather conditions exhibit distinct distributions in the high-dimensional feature space, hindering the performance of object detection networks. To address this issue, we propose a domain adaptation method based on adversarial learning. This approach ensures that extracted features have a similar distribution, even when input images originate from different data acquisition domains. Due to the lack of driving images recorded under a variety of weather conditions in existing datasets, we incorporate a semi-supervised learning framework to enhance detection performance by training with unlabeled images. Experimental results on public and our latest datasets demonstrate that the proposed adversarial learning technique surpasses recent traffic scene object detection networks across various driving scenarios. Code and datasets are available at https://github.com/daniel851218/all-weather-vehicle-detector.