Semi-supervised learning of pose-specific detector for human lying-pose detection

Daoxun Xia, Haojie Liu, Fang Guo, Weian Li

Abstract: Under the superlow-altitude aerial image, human lying-pose detection is an important problem in object detection. This paper is mainly focused on the application study of an Unmanned Aerial Vehicle (UAV) life detector after a disaster, and we study the problem of learning an effective pose-specific detector using weakly annotated images and a deep neural network. This typical approach (1) clusters a series of human poses for the human lying-pose and assigns an image-level label to all human lying-poses in each image and breaks them down into several categories; (2) trains multiple classifiers for each category using a deep neural network; and (3) uses the boosted semi-supervised CNN forest classifier to select a human lying-pose with high confidence scores as the positive instances for another round of training. Experiments on the XiaMen University Lying-Pose Dataset (XMULP) show that significant performance improvement can be achieved with our proposed method.