Residual Triplet Attention for Siamese Tracking

Abstract

Visual object tracking is a significant technique for various intelligent applications based on the Internet. Benefited by the application of attention mechanism, visual object tracking has made great progress. Recent popular attention mechanisms have been shown to be effective in improving the quality of the visual features, because attention mechanisms pay more attention to global information. However, most existing attention mechanisms applied in object tracking can only process the spatial or channel dimensions of feature maps independently, resulting in lack of information interaction among them. To address this issue, we propose a Siamese tracker based on our residual triplet attention. Firstly, we introduce a spatial attention module to improve the quality of the template and search region features. Secondly, we propose a residual triplet attention module (RTAM) suitable for object tracking. Feature maps have three dimensions: width, height, and channel. The first two contain spatial information, while the last one contains channel information. Treating each dimension of the feature maps equally, RTAM implements the information interaction between any two of the three dimensions simultaneously, which effectively improves the robustness and success rate of tracking. The extensive experiments on five benchmarks, including VOT2016, VOT2018, UAV123, OTB100, and GOT-10k, show that our proposed tracker achieves established performance.