Figure 2.
The overall of our Siamese anchor-free object tracking with multiscale spatial attention tracker, which consists of three modules: the Siamese-based subnetwork, the multiscale SAE block and the classification and regression subnetwork. The Siamese-based subnetwork (left side) utilizes the ResNet-50 as backbone to extract the feature of the last three stages for both the template branch and the search area branch. The backbone of these two branches shares the same structure. Those features are modified by the SAE block. The classification and regression subnetwork (right side), which takes the multiscale spatial attention features as input to predict the position of the target in search region. denotes the depth-wise convolution operation. denotes the channel-wise addition operation.