Pedestrian detection aided by scale-discriminative network

Zongqing Lu, Wenjian Zhang, Qingmin Liao

Abstract: Deep learning is greatly successful when used for pedestrian detection. However, we find that this method is barely satisfactory for multi-scale detection. Meanwhile, various solutions such as multi-scale classifiers have been developed (based on traditional methods) to handle this situation. Considering this, we propose a scale-discriminative classifier layer (SDC) that contains numerous classifiers to cope with different scales. To expand the capacity for small-scale pedestrian detection, we construct a full-scale layer that converges both high-level semantic features and low-level features. From the analysis above, a scale-discriminative network (SDN) for pedestrian detection was born. We apply this network to the Caltech pedestrian dataset, and the experimental results show that the SDN achieves state-of-the-art performance.