Abstract
Computational defogging using machine learning presents significant potential; however, its progress is hindered by the scarcity of large-scale datasets comprising real-world paired images with sufficiently dense fog. To address this limitation, we developed a binocular imaging system and introduced Stereofog—an open-source dataset comprising 10,067 paired clear and foggy images, with a majority captured under dense fog conditions. Utilizing this dataset, we trained a pix2pix image-to-image (I2I) translation model and achieved a complex wavelet structural similarity index (CW-SSIM) exceeding 0.7 and a peak signal-to-noise ratio (PSNR) above 17, specifically under dense fog conditions (characterized by a Laplacian variance, vL < 10). We note that Stereofog contains over 70% of dense-fog images. In contrast, models trained on synthetic data, or real-world images augmented with synthetic fog, exhibited suboptimal performance. Our comprehensive performance analysis highlights the model’s limitations, such as issues related to dataset diversity and hallucinations—challenges that are pervasive in machine-learning-based approaches. We also propose several strategies for future improvements. Our findings emphasize the promise of machine-learning techniques in computational defogging across diverse fog conditions. This work contributes to the field by offering a robust, open-source dataset that we anticipate will catalyze advancements in both algorithm development and data acquisition methodologies.