Learnable Laplacian Embedding Decomposition for One-Stream Visual Object Trackers
Omar Abdelaziz, Mahmoud Soliman, Mahmoud Alaa, Mohamed Shehata · 2025
Recent advancements in the one-stream tracking framework have demonstrated the power of unified feature learning and relation modelling for robust visual object tracking. The methods under this framework might suffer from noisy attention scores in the process of extraction of the feature maps within the transformer blocks. This paper introduces a novel Learnable Laplacian Embedding Decomposition (LLED) module designed to enhance one-stream trackers by explicitly decomposing the learned feature embeddings into a set of Laplacian embeddings, each representing a different scale of structural information that works as a band-pass filter after each encoder block. This filter rejects attention scores corresponding to extremely low or high frequencies, effectively suppressing noise and background distractors while emphasizing features most relevant to the target object. Experimental results on multiple challenging benchmarks, including GOT-10k, UAV123, and OTB2015, demonstrate that LLED, which is enhanced by the proposed band-pass filtering, significantly improves the localization accuracy of state-of-the-art one-stream trackers while maintaining their real-time speed. This work highlights the importance of structural information and its interplay with attentional mechanisms for precise tracking and provides a novel approach for integrating it into the efficient one-stream paradigm.