Similarity learning for template-based visual tracking

Most existing discriminative models for visual tracking are often formulated as supervised learning of a binary classification function, whose continuous output is then cast into a specific tracking framework as the confidence of the visual target. We argue that this might be less accurate since the classifier is learned for making binary decision, rather than predicting the similarity score between the candidate image patches and the true target. On the other hand, a generative tracker aims at learning a compact object representation for updating of the visual appearance. This, however, ignores the useful information from background regions surroundding the visual target, and hence might not well separate the visual target from the background distracters. We propose in this work a visual tracking scheme, in which a similarity function is explicitly learned in a generative tracking framework to significantly alleviate the drifting problem suffered by many existing trackers. Experimental results on various challenging human sequences, involving significant appearance changes, severe occlusions, and cluttered backgrounds, demonstrate the effectiveness of our approach compared to the state-of-the-art alternatives.