Adaptive Multi-scale Fusion Siamese Network for Visual Tracking
Xuechen Sun · 2022
The task of object tracking is to predict the size and position of a target in the subsequent frames of a video sequence, by giving the target's size and position in the initial frame. Target tracking has now become an important task in the field of computing and is widely used in numerous areas. Nowadays, deep based object tracking methods have outperformed traditional correlation filtering algorithms in terms of speed, where Siamese based trackers are not only simple in structure, better in performance, frame rate exceeding real time requirements but also very suitable for long-term tracking. In this paper, we expect to improve the performance of the Siamese network tracker by using multi-scale feature fusion, and to obtain a multi-scale feature map with complete semantic information and spatial-adaptive features by fusing deep and shallow network features through two mechanisms: weighted multi-scale fusion and adaptive multi-scale fusion modules. After training on the GOT-10K dataset and testing on the OTB2015 dataset, the experimental results demonstrate that the proposed SiamFC-ADF has significantly improved in terms of success and precision rate, with a 1.6% success rate and 2.1% precision rate improvement over SiamFC and the increase in computational effort and speed loss is negligible.