Abstract:
Visual object tracking aims to determine the state of a target object across video frames, starting
from its initial position in the first frame. Recently, one-stream Transformer-based trackers have
demonstrated impressive performance by jointly processing template and search region tokens
through attention mechanisms, allowing effective extraction of target-specific features. However,
the high proportion of background in the search region can weaken target representations and
reduce the tracker’s ability to accurately discriminate target from the background. To address
this limitation, we propose a Transformer-based tracking framework with a controlled attention
mechanism that suppresses irrelevant attention from background search tokens and enhances
target-focused feature learning. In early encoder layers of proposed tracker, attention from search
tokens to template tokens is blocked to prevent background interference, while template-to-search
attention is allowed for target-specific feature extraction. Following the early encoder layers, a
learnable module is introduced to estimate the probability of each search token belonging to
the target, using a multilayer perceptron. Based on the module’s predictions, search tokens
are divided into target and non-target groups, and attention from selected target tokens to the
template tokens is enabled to enhance target-focused feature learning. Experiments on GOT-10k
dataset demonstrate that our tracker outperforms recent state-of-the-art methods, achieving an
average overlap of 73.8%, a success rate of 84.1% at a threshold of 0.5, and a success rate of
70.1% at a threshold of 0.75, while maintaining real-time processing of 39 frames per second.
