Semantic part RCNN for real-world pedestrian detection

Recent advances in pedestrian detection, a fundamental problem in computer vision, have been attained by transferring the learned features of convolutional neural networks (CNN) to pedestrians. However, existing methods often show a significant drop in performance when heavy occlusion and deformation happen because most methods rely on holistic modeling. Unlike most previous deep models that directly learn a holistic detector, we introduce the semantic part information for learning the pedestrian detector. Rather than defining semantic parts manually, we detect key points of each pedestrian proposal and then extract six semantic parts according to the predicted key points, e.g., head, upper-body, left/right arms and legs. Then, we crop and resize the semantic parts and pad them with the original proposal images. The padded images containing semantic part information are passed through CNN for further classification. Extensive experiments demonstrate the effectiveness of adding semantic part information, which achieves superior performance on the Caltech benchmark dataset.