What Do Pedestrians See?: Visualizing Pedestrian-View Intersection Classification

Abstract

Extensive research has been carried out on intersection classification to assist the navigation in autonomous maneuvering of aerial, road, and cave mining vehicles. In contrast, our work tackles intersection classification at pedestrian-view level to support navigation of the slower and smaller robots for which it is too dangerous to steer on a normal road along with the usual vehicles. Particularly, we focus on investigating the kind of features a network may exploit in order to classify intersection at pedestrian-view. To this end, two sets of experiments have been conducted using an ImageNet-pretrained ResNet-18 architecture fine-tuned on our image-level pedestrian-view intersection classification dataset. First, ablation study is performed on layer depth to evaluate the importance of high-level feature, which demonstrated superiority in using all of the layers by yielding 77.56% accuracy. Second, to further clarify the need of such high level features, Class Activation Map (CAM) is applied to visualize the parts of an image that affect the most on a given prediction. The visualization justifies the high accuracy of an all-layers network.