.png){kind=icon}
.png){kind=icon}
Return
Learning orientational interaction-aware attention and localization refinement for object tracking
.png){kind=icon}
PRE
AI
DOI：10.1016/j.dsp.2024.104972.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Existing Siamese-based tracking methods have limitations on discriminative representation capabilities and matching guidance, making them insensitive to accurate target discrimination in complex scenarios. In this paper, we propose to learn orientational interaction-aware attention and localization refinement for object tracking. The interaction-aware information of features is mined in a learnable manner to achieve feature enhancement, and the joint refinement localization mechanism recalibrates the tracking state to maintain accurate target prediction. In particular, we first construct an orientational channel interaction-aware module to reweight channel maps with respect to modeled channel correlation relativity, thus selectively weakening the impact of irrelevant channels for the current task. Then we adopt the orientational spatial interaction-aware module to capture global contextual interactions. By aggregating long-range dependencies, the complementary beneficial semantics facilitate the enhancement of feature discrimination ability. Finally, we infer the tracking state of the target to determine whether to perform localization refinement. When ambiguity-guided states exist, we extract multiple peak regions to perform re-evaluation, providing accurate tracking results and suppression of potential distractors. The proposed algorithm is lightweight and portable, requiring significantly less training samples to achieve efficient training, thus exhibiting broad applicability in resource-constrained scenarios. Extensive experiments on several representative benchmarks demonstrate that the proposed method achieves competitive performance against state-of-the-art trackers.
Keywords:
Siamese network
Object tracking
Attention module
Localization refinement