arXiv new; foundation video world-model adaptation · P2 · 2026-07-03

RetailSMV：它和通用视觉自监督的关系在于：零售场景偏垂直，但比较 ego/exo 视角对 Cosmos3-Nano 视频 world model 适配的贡献，值得借鉴数据视角

中相关；详见方法、贡献和实验边界。

RetailSMV: Exocentric vs. Egocentric Adaptation of Foundation Video World Models in Retail　arXiv new; foundation video world-model adaptation　原文链接

编号2607.00310 优先级P2 类别arXiv new; foundation video world-model adaptation 会议arXiv new; foundation video world-model adaptation 方法零售场景偏垂直，但比较 ego/exo 视角对 Cosmos3-Nano 视频 world model 适配的贡献，值得借鉴数据视角来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：零售场景偏垂直，但比较 ego/exo 视角对 Cosmos3-Nano 视频 world model 适配的贡献，值得借鉴数据视角。中相关；详见方法、贡献和实验边界。

**Figureure 2 · Pipeline overview**Figure 2 Pipeline overview. The RetailSMV corpus (section 3.1) provides synchronized egocentric and exocentric video across five real-world supermarkets, yielding 16,120 ego and 15,985 exo clips. Three matched LoRA configurations of the pretrained Cosmos3-Nano foundation video model NVIDIA (2025) (egocentric-only, exocentric-only, and combined) are trained under identical hyperparameters and optimization budget, isolating training-data viewpoint as the variable of interest. Every configuration is evaluated on the same 200-clip stratified held-out test set under a paired statistical protocol.这张图概括 RetailSMV 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。

**Base vs**Base vs. RetailSMV-adapted video world model The same retail prompts are continued by a pretrained foundation model and by our adapted model. RetailSMV adaptation preserves scene layout, object permanence, and action grounding where the pretrained model drifts. Figure 1 Base vs. RetailSMV-adapted video world model. The same retail prompts are continued by the pretrained Cosmos3-Nano foundation model (top, red) and by our RetailSMV-adapted LoRA (bottom, green) under identical inference settings. RetailSMV adaptation preserves scene layout (hand-of watermelon), action grounding (weigh tomato crate), and physical geometry (open fridge) where the pretrained baseline drifts.这张可视化用来解释 RetailSMV 学到的中间表征或对齐关系。重点看它是否支持正文里的机制判断。

核心问题

它和通用视觉自监督的关系在于：零售场景偏垂直，但比较 ego/exo 视角对 Cosmos3-Nano 视频 world model 适配的贡献，值得借鉴数据视角。

方法拆解

零售场景偏垂直，但比较 ego/exo 视角对 Cosmos3-Nano 视频 world model 适配的贡献，值得借鉴数据视角

主要贡献

中相关；详见方法、贡献和实验边界。

实验看点

实验部分建议重点看两类证据：一是作者是否把方法收益和更强数据、更长训练、更大模型区分开；二是跨模型、跨数据或跨任务迁移是否还能保留同样趋势。

局限与读法

这篇论文的结论需要结合任务设置、训练数据规模和消融实验一起看；不要只凭单个指标判断它对通用视觉表征的价值。