Heng Wang

Attention mechanism (AM) is a widely used method for improving the performance of convolutional neural networks (CNNs) on computer vision tasks. Despite its pervasiveness, we have a poor understanding of what its effectiveness stems from. It is popularly believed that its effectiveness stems from the visual attention explanation, i.e., attention weights indicate the importance of feature and AM advocates focusing on the important part of an input image rather than ingesting the entire input. However, we find only a weak consistency exists between the attention weights of features and their importance. We verify the feature map multiplication that brings about high-order non-linearity into CNNs is crucial for the effectiveness of AM. Furthermore, we show an essential impact of feature map multiplication on the learned surfaces of CNNs. With the high-order non-linearity, feature map multiplication plays a regularization role on CNNs, which makes the learned curves smoother and more stable in-between real samples (test/training samples in datasets). Thus, compared to vanilla CNNs, CNNs equipped with AM are more robust to noises and yield smaller model sensitivity scores, which is the reason for their better performance.

Environment perception is essential for autonomous driving vehicles (ADVs), and vision plays a vital role in environmental perception. Object detection and semantic segmentation as the basic computer vision tasks have become the significant technology in the ADVs’ perception module. With the development of deep learning, the results of both tasks had an obvious leap. However, these advances have been driven by a powerful baseline system, which brings strict computing resource requirements. When deploying these two tasks on the same platform, real-time performance usually gets worse. In this paper, a method was presented for end-to-end lane segmentation and obstacle detection in real-time performance. In this method, a multi-task network was designed by fusing segmentation network architecture and detection network architecture. With a specific training strategy and a modified open dataset, this method has an excellent performance in detecting lane lines and obstacles simultaneously. When comparing with separately running detection and segmentation modules at the same computing platform, the method presented in this paper achieves better real-time performance and lower hardware requirements.