Scholarlink

<p>Real-time detection transformer (RT-DETR) has been proposed as a promising approach for real-time object detection. However, the performance of RT-DETR is limited due to: (a) the insufficient receptive fields of PResnet backbone, (b) lack of multi-level feature capabilities in Transformer encoder, and (c) deficient understanding of channel-wise information. To address these challenges, we introduce a wavelet guided real-time detection transformer with sparse attention (WS-RTDT) that integrates wavelet transform and sparse cubic attention mechanism for real-time object detection. Specifically, it first exploits a wavelet-guided convolution to expand the respective field of backbone to capture both high-frequency details and low-frequency contextual information. A sparse spatial encoder (SSE) is then proposed to learn feature maps of different scales and extract the multi-level features from images which employs a sparse cubic attention mechanism to process spatial and channel-wise feature correlation in an efficient manner. Experiment results demonstrate that WS-RTDT achieves the state-of-the-art performance in real-time object detection tasks, outperforming existing mainstream detectors in terms of efficiency and accuracy. Source code and pre-trained models are available at <a href="https://github.com/he13689/WSDETR">https://github.com/he13689/WSDETR</a>.</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

Wavelet guided real time detection transformer with sparse attention

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Yiqing He,
  • Zefeng Zheng,
  • Zhuowei Wang,
  • Hanwei Wu,
  • Yunyun Zhang,
  • Lianglun Cheng

摘要

Real-time detection transformer (RT-DETR) has been proposed as a promising approach for real-time object detection. However, the performance of RT-DETR is limited due to: (a) the insufficient receptive fields of PResnet backbone, (b) lack of multi-level feature capabilities in Transformer encoder, and (c) deficient understanding of channel-wise information. To address these challenges, we introduce a wavelet guided real-time detection transformer with sparse attention (WS-RTDT) that integrates wavelet transform and sparse cubic attention mechanism for real-time object detection. Specifically, it first exploits a wavelet-guided convolution to expand the respective field of backbone to capture both high-frequency details and low-frequency contextual information. A sparse spatial encoder (SSE) is then proposed to learn feature maps of different scales and extract the multi-level features from images which employs a sparse cubic attention mechanism to process spatial and channel-wise feature correlation in an efficient manner. Experiment results demonstrate that WS-RTDT achieves the state-of-the-art performance in real-time object detection tasks, outperforming existing mainstream detectors in terms of efficiency and accuracy. Source code and pre-trained models are available at https://github.com/he13689/WSDETR.