PulseAugur
实时 11:09:46
English(EN) SENTRY: SAM2-Enhanced Neighbor-Aware and Temporally Reasoned Memory for Visual Tracking

SENTRY模块通过时序一致性增强了基于SAM2的视觉跟踪

研究人员开发了SENTRY,一个旨在通过增强SAM2系统中的内存更新机制来改进视觉对象跟踪的新型模块。SENTRY通过在提交内存更新之前验证其时序一致性来解决遮挡或快速运动期间的漂移问题。这个无需训练、即插即用的模块会聚合分割假设,将它们回溯成短轨迹,并使用感知邻居的匹配来偏好时序和几何上一致的掩码。当集成到现有跟踪器中时,SENTRY在多个基准测试中展示了持续的性能提升,在多个数据集上取得了新的最先进成果,而无需改变基础架构。 AI

影响 通过稳定基于SAM2的系统中的内存更新,提高了视觉跟踪的准确性和鲁棒性。

排序理由 该集群包含一篇详细介绍视觉跟踪新方法的学术论文。

在 arXiv cs.CV 阅读 →

AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →

SENTRY模块通过时序一致性增强了基于SAM2的视觉跟踪

报道来源 [2]

  1. arXiv cs.CV TIER_1 English(EN) · Mohamad Alansari, Yonathan Michael, Hasan AlMarzouqi, Muzammal Naseer, Naoufel Werghi, Sajid Javed ·

    SENTRY: SAM2-Enhanced Neighbor-Aware and Temporally Reasoned Memory for Visual Tracking

    arXiv:2606.24449v1 Announce Type: new Abstract: We revisit the memory update mechanism in SAM2-based visual object tracking and identify confidence-only mask selection as the dominant cause of drift under occlusion, rapid motion, and distractors. We introduce SENTRY, a training-f…

  2. arXiv cs.CV TIER_1 English(EN) · Sajid Javed ·

    SENTRY: SAM2-Enhanced Neighbor-Aware and Temporally Reasoned Memory for Visual Tracking

    We revisit the memory update mechanism in SAM2-based visual object tracking and identify confidence-only mask selection as the dominant cause of drift under occlusion, rapid motion, and distractors. We introduce SENTRY, a training-free, plug-and-play, refine-before-write module t…