Scholarlink

<p>Transformer-based trackers have demonstrated impressive performance in visual object tracking due to their large receptive fields, which allow for early interaction between the template and search regions. However, their reliance on patch embedding often leads to significant compression of image content, potentially compromising the representation of spatial continuity, which may hinder tracker’s performance. To address the above issue, we propose LKTrack, a novel single-stream tracking framework based on large-kernel convolutional networks. This framework enables early interaction between the template and search regions through large kernels while ensuring image continuity through gradual downsampling convolutions. Moreover, The texture sensitivity of convolutional networks enhances object perception in challenging scenarios, especially those with cluttered backgrounds. Additionally, to compensate for the potential loss of localization accuracy caused by downsampling, the Upsampling Fusion (UF) Module that leverages shallow features for refining bounding box predictions is employed, thereby improving localization accuracy. As a purely convolution-based framework, our LKTrack achieves competitive results on TrackingNet (84.5% AUC) and LaSOT (71.2% AUC) and gains state-of-the-art performance on GOT-10k (74.1% AO) and UAV123 (71.3% AUC), surpassing most transformer-based trackers, demonstrating the effectiveness of large-kernel convolutional networks for the tracking task.</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

LKTrack: a novel tracking framework with large kernel network

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Hong Zhang,
  • Huakao Lin,
  • Ding Yuan,
  • Jianbo Song,
  • Hanyang Liu,
  • Yifan Yang

摘要

Transformer-based trackers have demonstrated impressive performance in visual object tracking due to their large receptive fields, which allow for early interaction between the template and search regions. However, their reliance on patch embedding often leads to significant compression of image content, potentially compromising the representation of spatial continuity, which may hinder tracker’s performance. To address the above issue, we propose LKTrack, a novel single-stream tracking framework based on large-kernel convolutional networks. This framework enables early interaction between the template and search regions through large kernels while ensuring image continuity through gradual downsampling convolutions. Moreover, The texture sensitivity of convolutional networks enhances object perception in challenging scenarios, especially those with cluttered backgrounds. Additionally, to compensate for the potential loss of localization accuracy caused by downsampling, the Upsampling Fusion (UF) Module that leverages shallow features for refining bounding box predictions is employed, thereby improving localization accuracy. As a purely convolution-based framework, our LKTrack achieves competitive results on TrackingNet (84.5% AUC) and LaSOT (71.2% AUC) and gains state-of-the-art performance on GOT-10k (74.1% AO) and UAV123 (71.3% AUC), surpassing most transformer-based trackers, demonstrating the effectiveness of large-kernel convolutional networks for the tracking task.