An Enhanced Visual Attention Siamese Network That Updates Template Features Online

Algorithm 2

Inference of the proposed framework.

 Input: test video; initial frame and bounding box of initial frame;

 Compute by the backbone network;

 Compute by the channel attention module;

 Compute by the spatial self-attention module;

 Preprocessing: crop and resize X and set three different scale patches .

 While test video is not empty do

  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;

  Upsampling feature map X to 272 × 272;

  Locate target center in feature map X by finding peak;

  Computer the offset of the upsampled map relative to the feature map;

  Computer the offset of the feature map relative to original image;

  Update target size and corresponding bounding box;

 end