Cascaded Regional Spatio-Temporal Feature-Routing Networks for Video Object Detection

This paper presents a cascaded regional spatiotemporal feature-routing networks for video object detection. Region proposal networks in faster region-based convolutional neural network (CNN) generate spatial proposals, whereas neglecting the temporal property of the videos. We incorporate the correlation filter tracking on the convolutional feature maps to explore an efficient and effective spatiotemporal region proposal generation method. To gradually refine the bounding boxes of proposals, three region classification and regression networks are cascaded. Feature maps from different layers in CNNs extract hierarchical information of the input, so we propose a router function which selects feature maps according to the scale of proposals. In addition, object co-occurrence inference is exploited to suppress conflicting false positives, which leads to a semantically coherent interpretation on the video. Extensive experiments on the Pascal VOC 2007 dataset and the ImageNet VID dataset show that the proposed method achieves the state-of-the-art performance for detecting unconstrained objects in cluttered scenes.