边缘视觉跟踪中VLM重定位的时机与净收益
Timing and Net Benefit of VLM Re-Localization in Edge Visual Tracking
张淏 1宋伟 1张铖阳 2罗志翔 2袁宇婷2
作者信息
- 1. 南昌航空大学航空宇航学院
- 2. 孝彭书院
- 折叠
摘要
[目的] 边缘视觉跟踪常在困难帧调用视觉语言模型（VLM）重定位。现有工作默认此类调用至多冗余，以调用次数度量其代价，而未回答某一次调用是否值得。[方法] 本文记录调用时刻的相对定位质量 ΔIoU = IoU(VLM 框, GT) − IoU(追踪器框, GT)。重定位以单帧观测替换追踪器已积累的状态，会清空其时序状态，因此 ΔIoU 为负时调用有害，为正时也未必有益。[结果] 该关系在 19 组配置上与净收益同向，三类对照下亦成立。据此本文给出两条规则：一条以两框 IoU 是否不低于 0.3 为判据接受 VLM 框，无需真值；另一条在序列开头做一次探测调用，由序列级门控决定整条序列是否继续调用，可把三组配置合计的净收益由 −0.102 提升到 +0.121，代价为一帧标注。[局限] 上述结论以 VLM 仅接收目标类别名与整帧图像、重定位为替换式、调用保持稀疏为前提。[结论] ΔIoU 的符号可离线判定单次调用是否有益。冷却周期以时间计的调度器在 1% 量级占空比下取得正收益，并在嵌入式边缘平台上复现。
Abstract
[Objective] Edge visual tracking calls a vision-language model (VLM) for re-localization on hard frames. Existing work treats such calls as at worst redundant and measures their cost by call count, without asking whether a single call is worthwhile.[Methods] This article records the relative localization quality at the moment of the call, IoU = IoU(VLM box, GT) IoU(tracker box, GT). Re-localization replaces the state a tracker has accumulated with a single-frame observation and clears its temporal state; a negative IoU therefore makes a call harmful, and a positive one does not guarantee a gain.[Results] The relation holds in sign across 19 configurations and under three controls. Two rules follow: accepting the VLM box only when the two-box IoU is at least 0.3, which needs no ground truth; and a probe call at the start of each sequence whose result decides, through a sequence-level gate, whether the rest of the sequence is called, lifting the summed net benefit of three configurations from 0.102 to +0.121 at the cost of one annotated frame.[Limitations] These conclusions presuppose a prompt that carries only the category name and the frame, replacement re-localization, and sparse invocation.[Conclusions] The sign of IoU gives an offline criterion for whether a call helps. A scheduler whose cooldown is set in time rather than in call budget attains a positive benefit at a duty cycle of about 1% and reproduces on an embedded edge platform.关键词
边缘计算/调用调度/视觉目标跟踪/视觉语言模型/相对定位质量Key words
Edge computing/Invocation scheduling/Visual object tracking/Vision-language models/Relative localization quality引用本文复制引用
张淏,宋伟,张铖阳,罗志翔,袁宇婷.边缘视觉跟踪中VLM重定位的时机与净收益[EB/OL].(2026-11-08)[2026-11-11].https://chinaxiv.org/abs/202610.00016.学科分类
计算技术、计算机技术
