SAF: Semantic Attention Fusion Mechanism for Pedestrian Detection

Benefiting from deep learning methods, pedestrian detection has witnessed a great progress in recent years. However, many pedestrian detectors are prone to detect background instances, especially under urban scenes, which results in plenty of false positive detections. In this paper, we propose a semantic attention fusion mechanism (SAF) to increase the discriminability of detector. The SAF includes two key components, attention modules and reverse fusion blocks. Different from previous attention mechanisms which use attention modules for re-weighting the top features of network directly, the outputs of our attention modules are fused by reverse fusion blocks from high level layers to low level layers step by step, which aims at generating strong semantic features for pedestrian detections. Experiments on CityPersons dataset demonstrate the effectiveness of our SAF.