Effects of Brightness and Class-Unbalanced Dataset on CNN Model Selection and Image Classification Considering Autonomous Driving

In addition to an approach of combining machine learning (ML) enhanced models and convolutional neural networks (CNNs) for adaptive CNN model selection, a thorough investigation study of the effects of 1) image brightness and 2) class-balanced/-unbalanced datasets is needed, considering image classification (and object detection) for autonomous driving in significantly different daytime and nighttime settings. In this empirical study, we comprehensively investigate the effects of these two main issues on CNN performance by using the ImageNet dataset, predictive models (premodel), and CNN models. Based on the experimental results and analysis, we reveal non-trivial pitfalls (up to 58% difference in top-1 accuracy in different class-balance datasets) and opportunities in classification accuracy by changing brightness levels and class-balance ratios in datasets.