SiamMLT: Siamese Hybrid Multi-layer Transformer Fusion Tracker

The current mainstream Siamese network cannot maximally discriminate between the target and the background because it cannot fully utilize the features extracted by a feature network. Here we propose a novel tracker called SiamMLT, which employs a convolutional neural network (CNN) as the backbone and transformer for multi-layer feature fusion. To fully exploit both the low-level and high-level features of the CNN network, transformer performs feature fusion to enhance the feature information and information-expression capabilities. Our tracker replaces the subsequent correlation operations with cross-attention, which generates a fusion vector of the template and search regions for subsequent classification and regression operation. Our method largely enhances the network’s ability to discriminate between objects and backgrounds and to perceive objects. The superior performance of our method is demonstrated in extensive experiments and evaluations on five challenging datasets. Finally, SiamMLT tracker achieves compelling performance compared with the state-of-the-art trackers.