Efficient Dual Attention Module for Real-Time Visual Tracking

Abstract

Attention mechanisms are of great potential to enhance discriminative capacity and adaptability of Convolutional Neural Networks. However, due to speed requirements of trackers, it is inappropriate to introduce attention modules that need high computational-cost into tracking algorithms. To employ attention mechanisms in tracking field, we propose Efficient Dual Attention Module (EDAM) which is a lightweight and effective module to extract attention in channel and spatial dimensions. Firstly, we aggregate spatial and channel information with self-attention mechanism by a simplified Non-local Network, which is called Attention-wise Global Pooling. Next, we extract dual attention by modeling inter-channel and inter-position relationships and finally fuse them in a sequential manner. With the use of dual attention, the network adaptively attends to discriminative and robust features of targets. Extensive experiments on a standard tracking benchmark demonstrate that our tracker with an EDAM embedded achieves favorable performance against state-of-the-art trackers while runs at real-time speed.