Spatial Remapped Network for Transformer Tracking
Kang Liu, Long Liu, Jiaqi Wang, Tong Li, Zhen Lin Wei · 2023
Recently, the transformer-based trackers have achieved significant performance improvements due to the powerful global modeling capabilities of attention. However, the existing trackers still suffer from the problem of effective discrimination between targets and similar interference when face with complex scenarios such as similar target distractor, occlusion, and significant target changes in the target neighborhood. In this work, we propose a spatial remapped network for transformer tracking, namely TrSRM. The spatial remapped (SRM)network establishes more significant discrepancy between targets and similar interference by remapping features, enhancing discriminative location ability to achieve more accurate target tracking. In addition, we design a training method based on Model-Agnostic Meta-Learning (MAML) for helping the SRM network to obtain a target tasksensitive initialization parameter. Experiments on benchmarks show that the proposed tracker outperforms several state-of-the-art methods, demonstrating the effectiveness of our method.