On Fast Trackers that are Robust to Partial Occlusions

Model-free tracking aims identify the location of particular objects or object parts in each frame of a video based on a single positive example. In our work, we (1) develop online-learning algorithms for part-based models that facilitate the use of these models in model-free tracking in order to improve robustness to partial occlusions, and (2) derive a probabilistic bound that facilitates rapid pruning of candidate locations in many popular trackers. Together with other recent advances in object detection and tracking, we believe these developments will ultimately contribute to solving the long-term tracking problem.