ACTION RECOGNITION IN STILL IMAGES BASED ON R-FCN DETECTOR

Still image action recognition is the task of localizing and classifying a person's action based on a single image. Most of the research on the field has focused on the action classification task, assuming that human bounding-box annotations are provided at both training and testing time. Thus, researchers have pursued different cues in order to characterize actions, often using multiple image patches. We argue that the action should be considered during the detection stage, considering that the existing state-of-the-art generic object detectors are not adequate for proposing candidate bounding boxes for action classification as they disregard human-object interactions. Our approach outperforms significantly methods based on general person detection by 16.5\% mean average precision (mAP) on the PPMI dataset and by 29.7\% on PASCAL VOC 2012. Our approach also achieves state-of-the-art results in both datasets, outperforming them by 1.2\% on the PPMI and by 3.7\% on PASCAL VOC 2012.