RGBT Tracking via Multimodal Bias Enhancement
Dengdi Sun, Gaotian Zhao, Zhuanlian Ding · 2025
RGBT tracking is mainly implemented by mutual enhancement and complementarity of two modalities, RGB and TIR, so that it can effectively perform target tracking in various complex environments. The multimodal interaction plays a critical part in it. Previous methods use symmetric operation in each interaction, which allows RGB and TIR to influence each other in each interaction. However, this approach ignores the problem that when a single modality is poor in quality, it may introduce background noise into the other modality. Moreover, in this coarse interaction, unnecessary redundant noise is actually introduced for the side with better modal quality. To overcome these limitations, we design a novel Multimodal Bias Enhancement (MBE) module, which is based on an adaptive token selection strategy that enables the better quality modality to perform cross-modal interactions on the poorer modality by recognizing the differences in the selection results. During each bias enhancement, the better modality performs a focused enhancement on the other modality, thus improving the features of the poorer modality and providing it with more robust multimodal context information. Our MBE module is inserted into the ViT backbone for biased modality enhancement to accomplish cross-modality interaction. We evaluate the MBE module on two RGBT tracking benchmarks, including RGBT234 and LasHeR. The experimental results show that our method achieves advanced performance.