Time Series Correlation for First-Person Videos

In this paper, an efficient feature encoding for firstperson video is introduced. The proposed method is appropriate for abstraction of high dimensional features such as those extracted from Convolutional Neural Networks (CNNs). The perframe extracted features are considered as time series, and the relations between them, in both temporal and spatial directions, are employed to represent the video descriptors. To find the relations, the time series are grouped and the linear correlation between each pair of groups are calculated. Furthermore, we split series in temporal direction in order to better focus on each local time window. The experiments show that our method outperforms previous methods such as Bag of Visual Word (BoVW), Improved Fisher Vector (IFV) and recently proposed Pooled Time Series (PoT) on the first-person DogCentric dataset. In addition, the presented method achieves a considerable improvement in computation time.