Abstract
RGB-T object tracking has advanced significantly recently, thanks to the emergence of numerous Transformer-based methods. However, although the conventional self-attention mechanism of transformers provides powerful global context modeling, its attention distribution can still be distracted by extensive background regions in challenging RGB-T scenarios, making it difficult to emphasize target-related local information consistently. To address this issue, we introduce novel, Importance Weighted Attention (IWA), which strengthens target-related attention interactions by suppressing redundant background responses during attention computation. To realise IWA, we propose a two-branch Siamese architecture based on the ViT backbone and design a bidirectional visible-thermal infrared modality interaction module. This module combines cross-attention and Importance Weighted Attention to facilitate information exchange between the RGB and TIR modalities. Thanks to IWA, the module focuses on the details of the target appearance, enhancing the quality of modeling the target region. The bidirectional cross-modal interaction promotes complementary feature fusion between RGB and TIR modalities as well, enabling more distinct target representation, while reducing background interference and improving tracking accuracy and robustness. Extensive experiments on the LasHeR, RGBT234, and RGBT210 datasets demonstrate that our approach outperforms existing methods, achieving state-of-the-art performance.