Feature Calibration Network for Occluded Pedestrian Detection

Pedestrian detection in the wild remains a challenging problem especially for scenes containing serious occlusion. In this paper, we propose novel feature learning method deep framework, referred to as Feature Calibration Network (FC-Net), adaptively detect pedestrians under various occlusions. FC-Net is based on observation that visible parts of are selective and decisive detection, implemented self-paced framework with self-activation (SA) module calibration (FC) module. new self-activated manner, learns features which highlight suppress occluded pedestrians. The SA estimates pedestrian activation maps by reusing classifier weights, without any additional parameter involved, therefore resulting an extremely parsimony model reinforce semantics features, while FC calibrates convolutional adaptive representation both pixel-wise region-based ways. Experiments CityPersons Caltech datasets demonstrate improves performance up 10% maintaining excellent non-occluded instances.