Multiple Stream Oriented Siamese Network for RGB-T Tracking
Yimo Wang, Сонглин Ду, Quan Zhou, Bin Kang · 2021 13th International Conference on Wireless Communications and Signal Processing (WCSP) · 2021
RGB-T tracker owns the capability of fusing two different yet complementary target observations, thus it will become a promising technology to fulfill all-weather tracking. Existing convolutional neural network based tracking methods often consider the multi-source oriented deep feature fusion from global viewpoint, leading to inevitable negative effect when feature maps of the target pair only contain little useful information. To solve this problem, we propose a four-stream oriented Siamese network, named as FS-Siamese, for RGB-T tracking. In particular, we introduce co-attention mechanism in bilinear pooling to explore the partial feature interaction between the RGB and thermal targets. This can effectively avoid uninformed image blocks disturbing feature embedding fusion. To enhance the efficiency of our Siamese network, we also propose an inner product based logistical loss for training the feature embedding and the graph convolutional neural network based bilinear pooling in an end-to-end manner. Extensive experiments on GTOT datasets demonstrate that the proposed method achieves state-of-the-art performance in the task of RGB-T tracking.