MEEM: Robust Tracking via Multiple Experts using Entropy Minimization

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.

An example is illustrated in the above figure. In (a), green rectangles show the results of our base tracker. After a period severe occlusion (e.g. frame #344), some previous tracker snapshot gives a different prediction at frame #374, shown in red. Thus, a disagreement among experts is detected, and we want to select the best expert to replace the current base tracker. The responses of the experts to the two predictions at frame #374 are displayed in (b) in the corresponding colors. Our multi-expert mechanism favors the snapshot at frame #250, which is less ambiguous when selecting between the red and green hypotheses, even though the current tracker gives the highest confidence score for the green prediction.

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.

Download

[PDF][Supplementary (Complete Experimental Results)]
[Data (10 newly collected sequences)] (1.9GB)
[Code]
[Precomputed Results on VTB1.0] **New**

Code Changelog:
04.24.2015: At some people's request, the precomputed results on VTB1.0 is made available.
10.03.2014: Removed the svmtrain_my.m and related files, which caused compatility issues in later versions (>=2013) of Matlab. The original purpose of using svmtrain_my.m was to suppress some solver errors (e.g. convergence-condition-not-met error) that can interrupt the tracking. You can still download the first version of the code at here, in case the later versions of svmtrain.m in Matlab affect the tracking results.
Users of our code are asked to cite the following publication:

@inproceedings{zhang2014meem,
title={{MEEM:} robust tracking via multiple experts using entropy minimization},
author={Zhang, Jianming and Ma, Shugao and Sclaroff, Stan},
booktitle={Proc. of the European Conference on Computer Vision (ECCV)},
year={2014}
}

DEMO


[latin](4.7M)

[ball] (6.0M)

[carRace](18.3M)

[boxing1](14.0M)

[dance](8.6M)

[boxing2](9.7M)

[ped1](2.3M)

[ped2](7.5M)

[billieJean](7.7M)

[rocky](3.4M)