Extended Joint Deep Learning for Pedestrian Detection

Abstract

In this paper, we propose an extended version of Unified Deep Network (UDN). The Extended UDN (EUDN) uses multiple deformation models that operate independently of each other and mixture of the responses of the models to estimate the detection label. The deformation models of the EUDN jointly learned in order to complement each other through penalized in-diversity loss measured from the average correlation between the models. In our experiments, we show that combining independently the deformation models (which are even if worse than existing one) reduces the error in the manner similar to the ensemble learning, and considering diversity of the individual models is more effective without considering diversity. Our approach is evaluated on the Caltech datasets and achieves better performance than the UDN.