.png){kind=icon}
.png){kind=icon}
Return
Dynamic language adjustment and asymmetric fusion for vision-language tracking
.png){kind=icon}
PRE
AI
C
M
Q
X
H
DOI：10.1007/s10586-026-06019-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Vision-language (VL) trackers improve tracking performance by introducing high-level semantic information of language, thus alleviating the limitations of a single visual modality. However, most existing VL trackers focus only on simple integration of visual and language features, ignoring that single-frame language descriptions become inaccurate as the target state changes dynamically. This inaccuracy affects the modeling of target semantic features and reduces tracking performance. To alleviate the above issue, this paper designs a novel VL tracking framework called DAVL-Track that achieves accurate modeling of the target by dynamically adjusting language information. Firstly, a real-time language adjustment module is proposed to improve the accuracy of language features. Based on historical context information, this module decomposes language features into three components and assigns weights based on the importance of each component to highlight target-related features and weaken target-irrelevant features. Secondly, to facilitate efficient VL interaction, a transformer-based asymmetric multi-modal fusion network is designed. This network takes full advantage of the multi-layer structure of the transformer encoder and adopts a unique asymmetric fusion strategy to gradually achieve deep fusion of visual and language features. Finally, a memory-based update module is introduced for acquiring high-quality historical context. It updates the context by selecting the most relevant and representative tracking results as memory templates to maintain high-quality temporal information. Experimental results on TNL2K, OTB-Lang, LaSOT and LaSOText datasets show that the proposed framework significantly improves the tracking accuracy and robustness.
Keywords:
Visual tracking
Vision-language tracking
Asymmetric multi-modal fusion
Language adjustment
Journal
C
IF:
0
Papers:
691
Citations:
1
Organization
Cited Papers