Scholarlink

<p>Online object detection is a crucial task related to the safety of autonomous driving during dynamic movement. While previous researches have achieved a significant breakthrough in low-latency detection by pursuing higher FPS, there has not been a clear analysis between detection speed and online object detection performance, particularly for non-real-time detectors (<i>e.g.</i>&#xa0;DETR-based detectors). To address this issue, we propose a metric named fAP to select a future frame as the matching target for the current frame according to the FPS of the detector. Through this new metric, we reduce the bias of detection performance evaluation caused by input frame inconsistency and verify the potential of the object detection model based on transformer architecture for the online object detection task. To this end, we develop a novel end-to-end object detector in conjunction with streaming perception, named Stream-DINO, which has better online object detection capability than other transformer-based models. Specifically, a Feature Encoder Acceleration Module is proposed to reduce the computation cost, thus make it focus on small target objects in each frame. We also introduce an Associated Frame Loss to utilize the object information captured in the associated frame to supervise the detection results and obtain accurate location of target objects. Notably, Stream-DINO is the first transformer-based online object detection method, and we achieve SOTA performance ( 19.9% in fAP and 18.6% in sAP) on Argoverse-HD dataset, outperforming the baseline (DINO) by 2.2% in fAP and 3.0% in sAP respectively.</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

Stream-DINO: exploring DETR-based online object detection with streaming perception

  • 期刊论文

  • 发表日期：

    2026年2月14日

查看全文
  • Man Zhang,
  • Yongqiang Zhang,
  • Rui Tian,
  • Yin Zhang,
  • Zian Zhang,
  • Jinwei Sun

摘要

Online object detection is a crucial task related to the safety of autonomous driving during dynamic movement. While previous researches have achieved a significant breakthrough in low-latency detection by pursuing higher FPS, there has not been a clear analysis between detection speed and online object detection performance, particularly for non-real-time detectors (e.g. DETR-based detectors). To address this issue, we propose a metric named fAP to select a future frame as the matching target for the current frame according to the FPS of the detector. Through this new metric, we reduce the bias of detection performance evaluation caused by input frame inconsistency and verify the potential of the object detection model based on transformer architecture for the online object detection task. To this end, we develop a novel end-to-end object detector in conjunction with streaming perception, named Stream-DINO, which has better online object detection capability than other transformer-based models. Specifically, a Feature Encoder Acceleration Module is proposed to reduce the computation cost, thus make it focus on small target objects in each frame. We also introduce an Associated Frame Loss to utilize the object information captured in the associated frame to supervise the detection results and obtain accurate location of target objects. Notably, Stream-DINO is the first transformer-based online object detection method, and we achieve SOTA performance ( 19.9% in fAP and 18.6% in sAP) on Argoverse-HD dataset, outperforming the baseline (DINO) by 2.2% in fAP and 3.0% in sAP respectively.