Focusing On Discrimination Between Road Conditions And Weather In Driving Video Analysis

We study an often ignored problem, the discrimination between road conditions and weather in driving videos, which may possibly lead to imperceptible errors on driving data analysis. We explore BDD100K, a common driving video database, and Kyushu Driving Data, a huge driving database created by ourselves. In our experiments, we use road condition labels and weather labels respectively to train several deep models on driving image sequences and demonstrate the difference between the two varieties of labels. The results indicate a significant difference between the two varieties, which leads to different performance of deep models.