An Improved Siamese Tracking Network Based On Self-Attention And Cross-Attention

Deep Siamese visual tracking network SiamRPN++ is found that its success rate and robustness is unsatisfactory when meeting complex scenes such as occlusion, large deformation, interference of similar objects and long-time tracking. Refer to these, we propose an improvement strategy based on self-attention and cross-attention mechanism. For backbone, we use Channel and Space self-attention modules, and we using different cross channel attention modules between template features and search features in every three RPN modules, finally using special self-attention on similarity feature maps. These tricks effectively suppress interference, improve the features' quality and make progress in robustness. Comparing with original SiamRPN++ with parameters from official open-source frame, PySOT, our network improves robustness of 3% on VOT2018, accuracy of 2% and success rate of 3% on OTB100.