Hierarchical Motion Excitation Network for Few-Shot Video Recognition

Most of the existing deep learning algorithms are supervised and rely on a tremendous number manually labeled samples. However, in most domains, due to scarcity samples or excessive cost labeling, it would be impracticable provide numerous training network. In this paper, few-shot video classification network termed Hierarchical Motion Excitation Network (HME-Net) is proposed from perspective accumulated feature-level motion information. An HME module composed (ME) Interval Frame (IFME) designed extract patterns adjacent frames interval frames. The can discover enhance motion-sensitive information original features. accumulative time window expanded four hierarchical manner, which achieves purpose increasing receptive field. After extensive experimentation, HME-Net demonstrated able consistently outperform models. On UCF101 HMDB51 datasets, our method established as new state-of-the-art technique for settings five-way three-shot five-shot recognition.