We propose a novel method for recognizing human actions from a single frame. Our approach utilizes the “action-ness” obtained by a pretrained model, enabling effective training for frame-wise action recognition without additional data. We also introduce several anomaly detection applications utilizing our method.