Figure 1. Two typical examples of detection results on the adverse
weather conditions adaptation experiments with different methods.
Semi-supervised method UBT [27] lacks awareness of objects in
the fog. Adversarial based GPA [49] attempts to exploit the objects
in the fog but gives some wrong predictions, such as the motorcy-
cle in the first row and the person in the second row. Our methods
can predict the boxes and categories more accurately.