Optical flow in deep visual tracking

Single-target tracking of generic objects is a difficult task since a trained tracker is given information present only in the first frame of a video. In recent years, increasingly many trackers have been based on deep neural networks that learn generic features relevant for tracking. This paper argues that deep architectures are often fit to learn implicit representations of optical flow. Optical flow is intuitively useful for tracking, but most deep trackers must learn it implicitly. This paper is among the first to study the role of optical flow in deep visual tracking. The architecture of a typical tracker is modified to reveal the presence of implicit representations of optical flow and to assess the effect of using the flow information more explicitly. The results show that the considered network learns implicitly an effective representation of optical flow. The implicit representation can be replaced by an explicit flow input without a notable effect on performance. Using the implicit and explicit representations at the same time does not improve tracking accuracy. The explicit flow input could allow constructing lighter networks for tracking.