An Enhanced Visual Attention Siamese Network That Updates Template Features Online

Algorithm 1

Offline training of the proposed framework.

	Input: random initialization the network parameters , golden threshold stochastic pooling T, spatial self-attention parameters of . Template Z and search patch X from GOT10K.
Preprocessing: crop and resize Z and X and set optimizer, loss function, and learning rate adjustment strategy.
While epoch > 0 and input video dataset is not empty do
Get template Z and corresponding bounding box;
Get search patch X and corresponding bounding box;
Compute , by the backbone network;
Compute , by the channel attention module;
Compute , by the spatial self-attention module;
Create sample positive and negative labels;
Compute and update template;
Computer response map of Z nad X;
Computer loss and update parameters;
Optimize loss to minimize.
end