稳健性(进化)
计算机科学
BitTorrent跟踪器
人工智能
计算机视觉
利用
钥匙(锁)
RGB颜色模型
跟踪(教育)
跟踪系统
模态(人机交互)
编码(集合论)
模板
代表(政治)
判别式
源代码
眼动
视频跟踪
航程(航空)
作者
Shihui Zhang,Junbin Su,Jiawei Zhang,Ziteng Xue,Zhipeng Zhang
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-06-30
卷期号:: 1-6
被引量:2
标识
DOI:10.1109/icme59968.2025.11209350
摘要
RGB-T tracking has shown great potential for tracking robustness owing to the assistance of thermal infrared information. Recent methods are designed to adapt pre-trained RGB trackers for RGB-T tracking, typically through lightweight modules for cross-modal interaction, known as parameter-efficient fine-tuning (PEFT). However, the limited number of tunable parameters typically restricts the model’s learning capacity, making it difficult to handle challenging scenarios. To address this limitation, we propose TGATrack, a template-guided framework designed for robust RGB-T tracking via low-rank adaption. The main idea of TGATrack is efficient target-centric modality complementation through two key components: 1) Template-guided Modality Mixer (TM-Mixer), which fully exploits the high-level target information in templates for target-aware representation alignment. 2) Dynamic Multimodal Combiner (DMC), which adaptively adjusts modality weights conditioned on the templates for joint prediction. Extensive experiments demonstrate that TGATrack achieves satisfactory results against state-of-the-art PEFT methods on popular RGB-T datasets with only 0.62M tunable parameters while exhibiting exceptional robustness in a wide range of challenging scenarios. The code is available at https://github.com/rp7sv/TGATrack.git.
科研通智能强力驱动
Strongly Powered by AbleSci AI