Free online tracking of an object given only the objects initial position and previous observations, within a tracking-by-detection framework
normally an object model is maintained via online updates which are intended to account for appearance changes of the target. However, the process of updating the model also brings the model drift problem, which is a key challenge in online visual tracking.
Model drift occurs because of factors like tracking failure, occlusions and misalignment of training samples can lead to bad model updates
Remedy→ incorporate the first frame template or prior knowledge in the online model update procedure
drawbacks of the remedy → relying on a fixed model prior tends to restrict the trackers ability to handle large object appearance changes.
Other remedy → use Censorship mechanism where an update is prevented when certain criteria are met(or not met). the detection of good or bad updates usually relies upon smoothness of assumptions for motion and appearance changes which are often violated in challenging scenarios. And once the censorship mechanism fails, these trackers will either miss the chance to evolve or get trapped in a background region due to the fact that the model can only evolve forward without a mechanism to correct for past mistakes.
instead of trying to prevent bad updates from happening we propose a formulation that can correct the effects of bad updates after they happen for this purpose we introduce a multi- expert tracking framework where a discriminative tracker and its former snapshots constitute an expert ensemble and the best expert is selected based on a minimum loss criterion to restore a tracker when a disagreement among the experts occurs Traditional loss functions which measure the discrepancy between the prediction and the true label are only applicable in supervised settings To get around this we propose a novel formulation of the tracking by detection problem so as to naturally introduce an entropy regularized optimization function as our expert selection criterion.
Bad model updates updates usually contaminate a trackers appearance model with inconsistent training samples thus leading to ambiguous inference. Ref fig1; During a period of severe occlusion the trackers online updates incorporate the wrong foreground image patch. After the patch reappears although the tracker is still responsive to the true target in red it starts to over fit the the wrong patch in green yielding an incorrect prediction. In contrast our formulation maintains a set of tracker snapshots throughout the tracking process. A past snapshot can be identified to localize the target wiht less ambiguity.
To implement the base tracker in our multi expert framework, we adopt an online SVM algorithm that approximates the offline version by employing compact prototype