Refocused Attention: Long Short-Term Rewards Guided Video Captioning

The adaptive cooperation of visual model and language model is essential for video captioning. However, due to the lack of proper guidance for each time step in end-to-end training, the over-dependence of language model often results in the invalidation of attention-based visual model, which is called 'Attention Defocus' problem in this paper. Based on an important observation that the recognition precision of entity word can reflect the effectiveness of the visual model, we propose a novel strategy called refocused attention to optimize the training and cooperating of visual model and language model, using ingenious guidance at appropriate time step. The strategy consists of a short-term-reward guided local entity recognition and a long-term-reward guided global relation understanding, neither requires any external training data. Moreover, a framework with hierarchical visual representations and hierarchical attention is established to fully exploit the potential strength of the proposed learning strategy. Extensive experiments demonstrate that the ingenious guidance strategy together with the optimized structure outperform state-of-the-art video captioning methods with relative improvements 7.7% in BLEU-4 and 5.0% in CIDEr-D on MSVD dataset, even without multi-modal features.