A Fine-Grained Spatial-Temporal Attention Model for Video Captioning

Abstract

Attention mechanism has been extensively used in video captioning tasks, which enables further development of deeper visual understanding. However, most existing video captioning methods apply the attention mechanism on the frame level, which only model the temporal structure and generated words, but ignore the region-level spatial information that provides accurate visual features corresponding to the semantic content. In this paper, we propose a fine-grained spatial-temporal attention model (FSTA), and the spatial information of objects appearing in the video will be our main concern. In the proposed FSTA, we achieve the spatial-hard attention at a fine-grained region level of objects through the mask pooling module and compute the temporal soft attention by using a two-layer LSTM network with attention mechanism to generate sentences. We test the proposed model on two benchmark datasets, namely, MSVD and MSR-VTT. The results indicate that our proposed FSTA model can achieve competitive performance against the state of the arts on both datasets.