Meta-Learning based Siamese Network with Channel-wise Self-attention for Visual Tracking
Rui Wang, Bin Kang, Wei‐Ping Zhu · 2021
A great deal of attention has been paid to Siamese networks in visual object tracking. However, the Siamese network often faces the problem of over-fitting when adapting to new scenes, where the training samples for scene adaptation are limited. In this paper, we propose a meta-learning based Siamese network with channel-wise self-attention (CSA) on the basis of SiamFC3s architecture to overcome this limitation, in which meta-learning mechanism is adopted for efficient scene adaptation and CSA is used for better representation of the target object. Specifically, in the meta-learning phase, only part of the neurons' parameters, called hyper parameters are updated, leading to lightweight adaptation, which not only reduces the computational burden of the network but also overcomes the over-fitting problem. Experimental results based on benchmark OTB-100 have shown that our meta-learning based SiamFC3s incorporating the proposed CSA outperforms the baseline method SiamFC3s by 3.4% in success rate and 5.6% in precision rate.