Automatic Video Captioning via Multi-channel Sequential Encoding

摘要

In this paper, we propose a novel two-stage video captioning framework composed of (1) a multi-channel video encoder and (2) a sentence-generating language decoder. Both of the encoder and decoder are based on recurrent neural networks with long-short-term-memory cells. Our system can take videos of arbitrary lengths as input. Compared with the previous sequence-to-sequence video captioning frameworks, the proposed model is able to handle multiple channels of video representations and jointly learn how to combine them. The proposed model is evaluated on two large-scale movie datasets (MPII Corpus and Montreal Video Description) and one YouTube dataset (Microsoft Video Description Corpus) and achieves the state-of-the-art performances. Furthermore, we extend the proposed model towards automatic American Sign Language recognition. To evaluate the performance of our model on this novel application, a new dataset for ASL video description is collected based on YouTube videos. Results on this dataset indicate that the proposed framework on ASL recognition is promising and will significantly benefit the independent communication between ASL users and others.