S2L: Single-Streamline For Complex Video Event Detection

In this paper, we focus on event detection in complex videos. Due to the large scale perspective changing and irregular camera moving, optical flow fails to capture accurate motion in these videos. To solve this problem, we propose a Single-StreamLine (S2L) model to implicitly represent motion information from two aspects. Specifically, 1) we build relationship between frames through temporal encoder for the fact that motion exists in consecutive video frames; 2) we combine low level appearance feature and high level semantic feature in the reason that motion can be observed via the changing in both appearance and semantics. Our experiments use YLI-MED dataset which is an open TRECVID-style video croups based on YFCC100M and includes ten video events. The visualization of feature shows the great power of our model and the experiment result improves at least 2% in accuracy compared with the current state-of-the-art.