The forth paper: LMCF Large Margin Object Tracking with Circulant
Feature Maps/ Author: Mengmeng Wang, Yong Liu, Zeyi Huang /
Publication information: CVPR2017
Outline: LMCF puts much effort in improving the tracking accuracy by a
multimodel detection method and an original high-confidence update
strategy, which is called APCE measure(average peak-to-correlation
energy measure). Firstly, they present structured-output SVM to obtain
the tracking region results. In order to alleviate the computation burden,
they utilize lots of math formulations to take the advantage of FFT and
correlation filters. After getting the response map in each frame, they
carefully select multiple peaks in the map, and deploy the large difference
between the highest and the other peaks to receive the detection result.
Surprisingly, APCE strategy can reflect the flunctuation of the feature
map to decide the model updating or not. There are more similar peaks,
the APCE value will be lower and the tracker ‘ model will not be updated.
Advantages: The tracker can perform well based on hand-crafted and
deep features. And the speed of it is really high. The careful mutimodel
detection and APCE strategy guarantee the good accuracy and robustness.
Disadvantages: The formulations which utilizes the CF and FFT into the
SO-SVM are sophisticated.
Feature Maps/ Author: Mengmeng Wang, Yong Liu, Zeyi Huang /
Publication information: CVPR2017
Outline: LMCF puts much effort in improving the tracking accuracy by a
multimodel detection method and an original high-confidence update
strategy, which is called APCE measure(average peak-to-correlation
energy measure). Firstly, they present structured-output SVM to obtain
the tracking region results. In order to alleviate the computation burden,
they utilize lots of math formulations to take the advantage of FFT and
correlation filters. After getting the response map in each frame, they
carefully select multiple peaks in the map, and deploy the large difference
between the highest and the other peaks to receive the detection result.
Surprisingly, APCE strategy can reflect the flunctuation of the feature
map to decide the model updating or not. There are more similar peaks,
the APCE value will be lower and the tracker ‘ model will not be updated.
Advantages: The tracker can perform well based on hand-crafted and
deep features. And the speed of it is really high. The careful mutimodel
detection and APCE strategy guarantee the good accuracy and robustness.
Disadvantages: The formulations which utilizes the CF and FFT into the
SO-SVM are sophisticated.