Pedestrian Detection with a Directly-Cascaded Deconvolution-Convolution Structure.

Driven by recent advances in deep learning, the accuracy of object detection has been tremendously improved. However, detecting small and blurred pedestrians still remains an open challenge. In this paper, we propose a novel neural network structure, which can be flexibly combined with powerful object detection systems for boosting pedestrian detection. The proposed structure contains two key modules: (i) a cascaded deconvolution-convolution (CDC) module to expand the resolution of feature maps, meanwhile, keep the crucial information in the feature maps; and (ii) a double-helix connection (DHC) module to effectively fuse shallow-level and deep-level features in the detection network. The CDC module enables the network to reuse features of the lower layers and learn richer features given low-resolution input. In addition, the DHC module incorporates the features learned in different layers in a novel and unified fashion. Extensive experiments on KITTI and Caltech Pedestrian datasets demonstrate that the proposed modules can be easily plugged into existing object detection networks (e. g., single-stage SSD and two-stage MSCNN) and consistently achieve better performance without bells and whistles.