We used a simulator named CARLA to collect some synthetic data, which are under different scenes and with different amounts of fog.
As shown in the figure, the images from left to right are as follows: the defogged image using our method, the reconstructed foggy image, and the ground truth foggy image. This shows that our approach can model both foggy and defogged scenes.
Additionally, we can control the amount of fog by manually adjusting the size of beta during the inference stage, which is beneficial for generating more foggy autonomous driving data.
