Transformer Tracker Based on Clipped SwinV2 Network
Na Li, Longyu Gu, Yun Gao · 2023
As Transformer technology continues to show its advantages in the field of computer vision, object tracking technology based on the Transformer network has also attracted much attention. In this paper, a Transformer object tracking algorithm based on an improved SwinV2 backbone network is proposed. First, the SwinV2 network is suitably trimmed to fit the object tracking application requirements by combining the requirements of feature dimensionality and computational speed of the object tracking application and retaining the advantages of the SwinV2 Transformer network in terms of the visual model. Then, the trimmed SwinV2 network is migrated to the SwinTrack base algorithm as the backbone network. Comparison experiments are conducted with the recent mainstream tracking algorithms based on three public test sets, LaSOT, GOT-10k, and OTB100, and the results show that the algorithm in this paper has better tracking performance.