MEEM: Robust Tracking via Multiple Experts Using Entropy Minimization

Abstract

Abstract. We propose a multi-expert restoration scheme to address the model drift problem in online tracking. In the proposed scheme, a tracker and its historical snapshots constitute an expert ensemble, where the best expert is selected to restore the current tracker when needed based on a minimum entropy criterion, so as to correct undesirable model updates. The base tracker in our formulation exploits an online SVM on a budget algorithm and an explicit feature mapping method for efficient model update and inference. In experiments, our tracking method achieves substantially better overall performance than 32 trackers on a benchmark dataset of 50 video sequences under various evaluation settings. In addition, in experiments with a newly collected dataset of challenging sequences, we show that the proposed multi-expert restoration scheme significantly improves the robustness of our base tracker, especially in scenarios with frequent occlusions and repetitive appearance variations.

Robust Visual Tracking by Exploiting the Historical Tracker Snapshots

“…Our work is most close to MEEM [29], but with significant differences summarized as follows. Firstly, in [29], the online SVM method is adopted as the base tracker, and the grid searching method is used to sample image patches.…”

“…Hong et al [14] adopt the hierarchical appearance model to track object through multilevel. In MEEM tracker [29], multiple experts are used to handle the model drift problem, which shows high tracking efficiency.…”

“…Firstly, in [29], the online SVM method is adopted as the base tracker, and the grid searching method is used to sample image patches. Our method introduces the ridge regression model to learn the temporal context correlation of the object rather than the binary classifier (online SVM).…”

“…Our method introduces the ridge regression model to learn the temporal context correlation of the object rather than the binary classifier (online SVM). Secondly, in [29], multiple experts are regarded independently, and the entropy based loss function is defined on the single expert. Furthermore, since our base tracker of correlation filter uses the regression model with dense sampling scheme, the response map shows much less ambiguity than the binary classifier.…”

“…Others use multiple experts working parallel to better discriminate the target from the background [15,24]. One of the representative work is that in [29], Zhang et al propose to use the multi-expert restoration scheme to address the model drift problem, where an entropy based loss function is defined to determine whether the current tracker is reliable and should be restored to the historical tracker. The proposed tracking framework adopts online SVM as the base tracker, which shows very robust performance.…”