Transformer Union Convolution Network for visual object tracking

Abstract

In visual object tracking, the existing Siamese-like neural network utilizes the superposition of different convolution modules to extract feature maps, which makes the network unable to obtain sufficient global information. Inspired by Transformer, we present a novel method named Transformer Union Convolution Network (TUC-Net) to take the advantage of its encoder–decoder structure. Through the combination of attention mechanism and convolution, the interaction of local and global information between image patches is realized. Specifically, the proposed method includes a feature interaction stage based on feature perception head followed by encoder and a feature fusion stage based on U-Net-like convolution structure and decoder. Experiments shows that our TUC-Net achieves promising results on five challenging datasets compared with state-of-the-art methods. The average running speed of our tracker is approximatively 20 fps on GPU. Besides, we design a dual-resolution camera system with high resolution and large field of view. TUC-Net can be well-equipped with the system and get excellent tracking results.