Adaptive Augmentation of Imbalanced Class Distribution in Road Segmentation

Deep learning has achieved significant improvements in various tasks in Computer Vision. However, acquiring a large number of the dataset is a challenge in real-world applications, especially if they are new class objects for Deep Learning. Furthermore, the distribution of classes in the dataset is often imbalanced - a bottleneck of the neural network’s performance in classification. One possible real-world application is road segmentation, which is crucial for autonomous driving and sophisticated driver assistance systems to comprehend the driving environment. Recent years have seen significant advancements in road segmentation with the help of Deep Learning. Inaccurate road boundaries and lighting fluctuations such as shadows and overexposed zones are still challenging issues. Prediction performance is also impacted by an improper class distribution, which arises because most image pixels belong to the background (negative class), while the goal is to identify road pixels (positive class). In this paper, we focus on the topic of "visual road classification," where the target is to label each pixel as containing either a road or a background. We tackle this task by implementing a novel Adaptive Augmentation algorithm by integrating with some recently suggested encoder-decoder based convolutional neural network architecture and compare the qualitative and quantitative experimental results with traditional augmentation algorithm. The proposed method uses an adaptive augmentation module to improve performance under improper class distribution conditions. Experimental results show that the suggested method achieves higher segmentation accuracy than state-of-the-art methods on the KITTI road detection benchmark datasets.