相对定位质量与 VLM 重锚净收益的符号：一项边缘视觉跟踪的实证研究
Relative localization quality and the sign of VLM re-anchoring benefit: an empirical study in edge visual tracking
摘要
[目的] 边缘视觉跟踪在困难帧调用视觉语言模型（VLM），其收益是否为正尚无定论。[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)。[结果] 19 组配置中 ΔIoU 为负者净收益均为负、为正者未必为正，故 ΔIoU 是必要而非充分条件；目标级门控（需一帧标注或人工校验）使三组配置净收益之和由 -0.102 升至 +0.121。[局限] 该结论以稀疏调用、单一 3B VLM、替换式重锚设定，以及 VLM 仅接收类别名与整帧图像为前提。[结论] ΔIoU 的符号可离线判定单次调用是否有益；由此得到的免真值规则——仅在两框 IoU 不低于 0.3 时接受 VLM 框——可把两处有害配置的净收益由负转正，代价是有益配置让出约六成收益。
Abstract
[Objective] Edge visual tracking calls a vision-language model (VLM) to re-localize on hard frames, but whether a single call can hurt is unknown.[Methods] This article records the relative localization quality at the moment of the call, IoU = IoU(VLM box, GT) IoU(tracker box, GT).[Results] IoU is necessary but not sufficient for a gain: all 19 configurations with a negative IoU lose; a gate that costs one annotated frame lifts the net benefit from 0.102 to +0.121.[Limitations] These conclusions presuppose sparse invocation, a single 3B VLM, replacement re-anchoring, and a prompt that carries only the category name and the frame.[Conclusions] The sign of IoU gives an offline criterion for whether a call helps; the ground-truth-free rule it yields accepting the VLM box only when the two-box IoU is at least 0.3 turns both harmful configurations positive, at the cost of about sixty percent of the benefit on the beneficial ones.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.相对定位质量与 VLM 重锚净收益的符号：一项边缘视觉跟踪的实证研究[EB/OL].(2026-10-05)[2026-10-12].https://chinaxiv.org/abs/202610.00016.学科分类
工程基础科学
