DPTracker : Dynamic prompter for RGB-D tracking

Zhao, J. and Su, J. and Liu, Y. and Liu, J. and Wang, M. (2026) DPTracker : Dynamic prompter for RGB-D tracking. Pattern Recognition Letters, 204. pp. 79-85. ISSN 0167-8655

Text (PR_Letters_clean)
PR_Letters_clean.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (1MB)

Abstract

RGB-D object tracking aims to accurately localize a target across video frames by jointly leveraging RGB and depth modalities. However, existing RGB-D trackers often suffer from limited adaptability to varying scene conditions and inconsistent reliability of modalities. To address these challenges, we propose DPTracker (Dynamic Prompt Tracker), a novel RGB-D tracking framework that leverages visual prompt learning and dynamic modality prompting. Specifically, DPTracker introduces a Modality Effectiveness Predictor (MEP) to estimate the validity of each modality, and a Dynamic Prompter (DP) that adaptively adjusts the fusion intensity of RGB and depth information during prompting. By dynamically reweighting modal contributions, DPTracker effectively suppresses low-quality depth cues while enhancing reliable RGB information. Built upon a frozen pre-trained RGB tracker, the proposed framework only fine-tunes lightweight prompt-related parameters, substantially reducing computational cost and data requirements. Extensive experiments on three benchmark datasets, including DepthTrack, CDTB, and VOT-RGBD2022, demonstrate that DPTracker achieves state-of-the-art performance with superior robustness and generalization capability, thereby validating the effectiveness of dynamic modality prompting in RGB-D tracking. Our code is available at: https://github.com/neilzhao996/DPTracker.