Siamese Region Proposal Networks and Attention Module for Real-time Visual Tracking

Dong，Hang; Zeng，Yuan; Gong，Yi

10.1145/3447450.3447475

2020-12-25

ACM International Conference Proceeding Series

PartF168342

154-160

Recently, the region proposal networks have been combined with the Siamese networks for visualtracking and have achieved great attention due to their balanced accuracy and speed. However, it isstill challenging to track object with simultaneous requirements on robustness and discrimination power. A key to balance the online visual tracking accuracy and speed is to learn abundant features.In this paper, we propose an attention module based Siamese region proposal networks, named AM-Siam, for real-time visual tracking task. The basic idea is to refine neural features extracted from convolutional neural networks with channel attention module and spatial attention module, and providereliable visual attentional features for tracking. In addition, we design a multi- task loss function with balanced L1 loss to accelerate convergence speed of the proposed tracking network. The proposed AM-Siam is trained off-line in an end-to-end pattern and does not update the network parameters during tracking. Experiments on the Visual Object Tracking (VOT) show that AM-Siam achieves state-of-the-art results with tracking accuracy of 61% and tracking speed of 189 fps on the VOT. Moreover, experimental results demonstrate the effectiveness of our proposed AM-Siam tracker compared withstate-of-the-art trackers.

Attention module Real-time visual tracking Siamese networks

第一

英语

EI

20211610228973

Convolutional neural networks ; Image processing ; Video signal processing

Television Systems and Equipment:716.4

2-s2.0-85104177137

Scopus

Department of Electronics and Electrical Engineering,Southern University of Science and Technology (SUSTech),Shenzhen,518055,China

Dong，Hang,Zeng，Yuan,Gong，Yi. Siamese Region Proposal Networks and Attention Module for Real-time Visual Tracking[C],2020:154-160.