Fast cascaded action localization in video using frame alignment

Locating human actions in videos is challenging because of the complexity and variability of human motions, as well as of the amount of video data to be searched. We propose a method that detects and locates a set of actions in a video database by taking into account their temporal structure at the frame level. While other methods aggregate frames into action parts, we leverage the complementarity between aggregation and frame level comparison of sequences. Combining these two techniques in a cascade, we aim to address large scale retrieval. Evaluation on popular datasets show state of the art results, as well as efficient detection and low storage requirements.