Efficiently Handling Scale Variation for Pedestrian Detection

摘要

Pedestrian detection is a popular yet challenging research topic in the computer vision community. Although it has achieved great progress in recent years, it still remains an open question how to handle scale variation, which commonly exists in real world applications. To address this problem, this paper presents a novel pedestrian detector to better classify and regress proposals of different scales given by a region proposal network (RPN). Specifically, we have made the following major modifications to the Adapted FasterRCNN baseline. First, we divide all proposals into small and large pools according to their scales, and deal with each pool in a separate classification network. Also, we employ two auxiliary supervisions to balance the effect of two parts of proposals on the back propagation. It is worth noting that the proposed new detector does not bring extra computational overhead and only introduces very few additional parameters. We have conducted experiments on the CityPersons, Caltech and ETH datasets and achieved significant improvements to the baseline method, especially on the small scale subset. In particular, on the CityPersons and ETH datasets, our method surpasses previous state-of-the-art methods with lower computational costs at test time.