Low Visibility Street Scenes Recognition with Augmented Image Sets

摘要

Recently, deep learning is one of the popular ways of object detection, and the most important thing in object detection is the dataset, which is used in the training process. However, when we look over the dataset, we can find that most of accessible dataset is taken during the good weather (e.g., sunny day) with the simple environmen, most of the testing pictures and videos are also taken in the same environment. When we turn to test the model that is trained by bright and simple dataset during the bad weather (e.g., rainy day, cloudy day stormy, foggy day) or in the dim place (e.g., tunnel, night, dusk, under the object's shade), we will get a bad performance. To solve this problem, we use Photoshop to modify the original dataset, try to make the dataset more versatile and larger in number of images according to the real environment. Our experiments show that we improve the accuracy from 98.6% to 99.3% for the daytime testing dataset. And for the nighttime testing dataset, we improve the accuracy from 22% to 50%.