Combining simple trackers using structural SVMs for offline single object tracking

We introduce MMTrack, our single-target tracking system, that combines cluster-based and adaptive appearance modeling. MMTrack uses SVMs for aggregating simple trackers. We focus on modeling tracking as a structured output prediction task where the goal is to find a sequence of interdependent locations of the target given a video. Since "bad" trajectories are usually not given, we require a principled way for automatically generating them. Following recent advances in machine learning, we discriminatively learn the tracking task by first generating "bad" trajectories and then employing a max-margin criterion to distinguish among ground truth trajectories and all other possibilities. Our framework for tracking can be of general interest since one can add or remove trackers easily to obtain a desired tracker. Our method enjoys robustness against occlusion, drift and appearance change. We applied our framework to single pedestrian tracking and experimentally demonstrated the effectiveness of our method on a real-world dataset.