Object Tracking Paper(5)--Understanding and Diagnosing Visual Tracking Systems

The fifth Paper: Understanding and Diagnosing Visual Tracking
Systems/ Author: Naiyan Wang, Jianping Shi, Dit-Yan Yeung, Jiaya Jia/
Publication information: ICCV 2015
Outline: This paper about tracking aims to break the DCF trackers into
five constituent components, namely, motion model, feature extractor,
observation model, model updater, ensemble post-processor. They control
the variation of one part, and conduct the experiment to find the key
factor in visual object tracking. They conclude that the feature extractor
plays the most important part in tracking. And when the features is
powerful enough, such as HOG, the difference between SVM and ridge
regression is very small. The model updating scheme can impact the
tracker’s performance significantly. The Excessive updating frequency
may cause drift problem and let the model contaminated by bad samples.
And an effective ensemble processor can improve the performance. The
diversity of the trackers could promote the robustness and accuracy.
Advantages: This conclusion is reliable and reasonable. They tell the
researchers what should be put much effort on.
Disadvantages: The features they discuss is hand-crafted and deep
features are not included. The trackers speed should be considered.