Siamese Network for Object Tracking with Diffusion Model
Jiacheng Zhang, Yifeng Zhang · 2023
Recently, Siamese networks have drawn great attention in object tracking community because of their balanced accuracy and speed. However, due to the scarcity of target samples in the training period, Siamese networks suffer from low adaptability to the variations of target appearance in realistic tracking scenarios. Therefore, Siamese network for object tracking with diffusion model is proposed. Based on the target template labeled with bounding box, the network generates more target samples with ADM-G, and compares the feature map of samples to the ones of images from the searching areas in the classic SiamRPN++ architecture. High quality and diversity samples are highly correlated with the object, which greatly enrich the object sample, and enhance the quality of the feature extraction. As demonstrated by experiments, the Diff-SiamRPN++ tracker is superior to the others on different benchmarks.