An Exploration of Why and When Pedestrian Detection Fails

This paper undergoes a finer-grained analysis of current state-of-the-art in pedestrian detection, with the aims of discovering insights into why and when detection fails. Current pedestrian detection research studies are often measured and compared by a single summarizing metric across datasets. The progress in the field is measured by comparing the metric over the years for a given dataset. Nonetheless, this type of analysis may hinder development by ignoring the strengths and limitations of each method as well as the role of dataset-specific characteristics. For the experiments we employ two pedestrian detection datasets, Caltech and KITTI, and highlight their differences. The datasets are used in order to understand in what ways methods fail, and the impact of attributes, occlusion, and other challenges. Finally, the analysis is used to identify promising next steps for researchers.