Scholarlink

<p>Convolutional Neural Networks (CNNs) and Transformers have achieved significant progress in visual tracking by learning powerful feature representations. However, existing attention-based trackers still face two critical limitations: (1) they struggle to capture fine-grained similarity differences between features, leading to poor target-background discrimination in cluttered scenes; (2) they often lack effective multi-scale interaction, which limits robustness under large-scale variations. To address these issues, we propose a novel tracking framework. The framework consists of three main components: first, a feature extraction network based on the Light Multi-Scale Fusion Convolution Module (LMFC) and the Cross-Scale Parallel Attention Module (CSPA); second, a feature fusion module based on the Global Perception Attention Module (GPAtten); and third, a tracking prediction head. Specifically, the LMFC employs grouped convolutions with varying dilation rates to capture multi-scale representations efficiently, while the CSPA enhances cross-scale feature interaction in a parallel manner, improving feature richness and discrimination. The GPAtten then encodes fine-grained features with contextual information to further strengthen foreground representation. Extensive experiments on six large-scale benchmarks (LaSOT, TrackingNet, NFS, UAV123, GOT-10K, and TNL2K) demonstrate that our tracker consistently outperforms state-of-the-art methods. Specifically, on LaSOT, our method achieves a success rate (AUC) of 68.4%, outperforming ABTrack-AViT by 5%; on GOT-10k, it attains an average overlap (AO) of 72.4%, surpassing EMAT by 3.7%. Moreover, with only 52.2M parameters, our tracker runs at 107 FPS on a single GPU, satisfying real-time requirements without compromising accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Learning Global Perception and Cross-Scale Dependencies for Visual Tracking

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Wenshuang Zhang,
  • Lu Li,
  • Jun Wang

摘要

Convolutional Neural Networks (CNNs) and Transformers have achieved significant progress in visual tracking by learning powerful feature representations. However, existing attention-based trackers still face two critical limitations: (1) they struggle to capture fine-grained similarity differences between features, leading to poor target-background discrimination in cluttered scenes; (2) they often lack effective multi-scale interaction, which limits robustness under large-scale variations. To address these issues, we propose a novel tracking framework. The framework consists of three main components: first, a feature extraction network based on the Light Multi-Scale Fusion Convolution Module (LMFC) and the Cross-Scale Parallel Attention Module (CSPA); second, a feature fusion module based on the Global Perception Attention Module (GPAtten); and third, a tracking prediction head. Specifically, the LMFC employs grouped convolutions with varying dilation rates to capture multi-scale representations efficiently, while the CSPA enhances cross-scale feature interaction in a parallel manner, improving feature richness and discrimination. The GPAtten then encodes fine-grained features with contextual information to further strengthen foreground representation. Extensive experiments on six large-scale benchmarks (LaSOT, TrackingNet, NFS, UAV123, GOT-10K, and TNL2K) demonstrate that our tracker consistently outperforms state-of-the-art methods. Specifically, on LaSOT, our method achieves a success rate (AUC) of 68.4%, outperforming ABTrack-AViT by 5%; on GOT-10k, it attains an average overlap (AO) of 72.4%, surpassing EMAT by 3.7%. Moreover, with only 52.2M parameters, our tracker runs at 107 FPS on a single GPU, satisfying real-time requirements without compromising accuracy.