Video Summarization with Encoder-Decoder Based Graph-Attention Networks

Video summarization aims to generate a brief summary containing key video information by eliminating redundant information in videos. Video summarization faces many challenges such as the high cost of video data annotation, the lack of professional datasets, and the inaccuracy of keyframes selection algorithms. To address the above issues, we propose a non-manual dataset annotation method based on the number of online video barrage, and construct a new video summarization dataset based on this method. In addition, we propose an encoder-decoder based graph-attention video summarization algorithm, which introduces a graph attention mechanism and only assigns different weights to relevant sampled frame features to improve the precision of keyframes selection. Our method achieved state-of-the-art results on both benchmark datasets (SumMe and TVsum) and our new dataset.