Structure Preserving Object Tracking

Model-free trackers can track arbitrary objects based on a single (bounding-box) annotation of the object. Whilst the performance of model-free trackers has recently improved significantly, simultaneously tracking multiple objects with similar appearance remains very hard. In this paper, we propose a new multi-object model-free tracker (based on tracking-by-detection) that resolves this problem by incorporating spatial constraints between the objects. The spatial constraints are learned along with the object detectors using an online structured SVM algorithm. The experimental evaluation of our structure-preserving object tracker (SPOT) reveals significant performance improvements in multi-object tracking. We also show that SPOT can improve the performance of single-object trackers by simultaneously tracking different parts of the object.

1 Introduction

Object tracking is a fundamental problem in computer vision with applications in a wide range of domains. Whereas significant progress has been made in tracking specific objects (e.g., faces [22], humans [11], and rigid objects [15]), tracking generic objects remains hard. Since manually annotating sufficient examples of all objects in the world is prohibitively expensive and time-consuming, recently, approaches for model-free tracking have received increased interest [2], [12]. In model-free tracking, the object of interest is manually annotated in the first frame of a video sequence (using a rectangular bounding box). The annotated object needs to be tracked throughout the remainder of the video. Model-free tracking is a challenging task because (1) little information is available about the object to be tracked, (2) this information is ambiguous in the sense that the initial bounding box only approximately distinguishes the object of interest from the background, and (3) the object appearance may change drastically over time.