基于主题迁移的跨视角动作识别

Cross-view Action Recognition by Topic Transfer

利用迁移学习的思想,提出了一个主题迁移模型(topic transfer model)用于跨视角的动作识别。借助源视角视频和目标视角视频,学习一个迁移模型,利用这个模型来实现对目标视角下视频的分类。具体方法是在源视角下训练一个主题模型,将反应源视角的语义信息传递到目标视角中,然后在目标视角中训练一个主题模型,实现跨视角的动作表示,利用支持向量机进行动作的训练和分类。实验结果验证了该方法的有效性。

A novel approach is proposed to recognizing human actions from different views by view knowledge transfer. A topic model was trained in the source view, the semantic topics were transferred to the target view for learning a new topic model via a semi-latent manner. For a new video in target view, the topics inferenced by the target topic model can be transferred to the source view for action recognition.