[target tracking] | template update time context information (updatenet) "learning the model update for Siamese trackers"

Article title ：《Learning the Model Update for Siamese Trackers》
Article address ：[1] arXiv,[2] ICCV2019 openaccess
github Address ：https://github.com/zhanglichao/updatenet

Update strategy

Use only the initial template of the first frame

Early Siamese In the tracker , Initialize the target template in the first frame , This template is used for tracking later . However, in the actual task, the change of target appearance is usually obvious , If the target template cannot be updated effectively , Soon the tracking will fail .

Linear update

To solve this problem , At present, some people have made some simple update strategies , Through the moving average method of fixed learning rate , Update the target template linearly ：

We think this method has some disadvantages ：

• For all the videos , It uses a fixed update rate . But in different cases , The need for updates is different , The factors to consider are complex . For example, camera motion blur , Dynamic changes at different times, etc .
• This update rate applies to all spatial dimensions of the template （ Including channel dimension ） It's also fixed on . But sometimes we only need to update a part of the template , For example, when occlusion occurs .
• The tracker cannot recover from the cumulative drift . As the template is updated , We lost T 0 The original information of , And the template of the first frame T0 That's what we first wanted to track .
• The formula （1） Just a very simple linear combination , But we need a flexible update mechanism , It's necessary to change the appearance when the goal is very complex . It may be better to use more complex combinatorial functions .

Adaptive update

The update algorithm will always consider the template of the first frame Provides highly reliable information , It enhances the robustness to the later tracking drift .

ref
translate
https://blog.csdn.net/PAN_Andy/article/details/103008079
https://blog.csdn.net/Yemiekai/article/details/123014839、