LLLMWIKI
正文元数据首页

来源笔记 · 更新于 2026-07-14

Swin Transformer移位窗口与层级视觉 Transformer

#near-cvpr-2025#vision-foundations#representation-learning#primary-source

Swin Transformer：层级视觉 Transformer

一句话结论

Swin 用移位局部窗口和层级特征金字塔，把 Transformer 扩展为同时适合分类、检测与分割的通用视觉主干。

问题定义

怎样避免全局自注意力对高分辨率图像的平方复杂度，并产生密集视觉任务所需的多尺度表示？

方法概述

窗口内注意力控制计算，交替移位窗口传递跨窗信息，patch merging 形成逐级降低分辨率的层级表示。

关键发现

  • 在分类、检测、实例分割与语义分割上验证同一主干。
  • 局部计算与层级输出是适配高分辨率密集任务的核心。
  • Transformer 与 CNN 的边界从“有没有卷积”转向计算结构与归纳偏置。

局限或疑问

  • FLOPs 不直接等于硬件延迟。
  • 局部窗口的全局关系需要跨层传播。
  • 强结果不能隔离训练配方与架构单因素。

原始链接

相关页面

元数据
{
  "id": "2026-07-14-swin-transformer",
  "type": "source",
  "title": "Swin Transformer：移位窗口与层级视觉 Transformer",
  "status": "reviewed",
  "created": "2026-07-14",
  "updated": "2026-07-14",
  "venue": "ICCV 2021",
  "ingested_at": "2026-07-14",
  "tags": [
    "near-cvpr-2025",
    "vision-foundations",
    "representation-learning",
    "primary-source"
  ],
  "note_status": "reviewed",
  "source_type": "paper",
  "authors": [
    "Ze Liu",
    "Yutong Lin",
    "Yue Cao",
    "Han Hu",
    "Yixuan Wei",
    "Zheng Zhang",
    "Stephen Lin",
    "Baining Guo"
  ],
  "published_at": "2021-01-01",
  "canonical_links": [
    "https://arxiv.org/abs/2103.14030",
    "https://openaccess.thecvf.com/content/ICCV2021/html/Liu_Swin_Transformer_Hierarchical_Vision_Transformer_Using_Shifted_Windows_ICCV_2021_paper.html",
    "https://arxiv.org/pdf/2103.14030",
    "https://github.com/microsoft/Swin-Transformer"
  ],
  "raw_entry": "raw/ingest/2026-07-14-swin-transformer/",
  "analysis_note": "raw/ingest/2026-07-14-swin-transformer/analysis.md",
  "topics": [
    "topics/vision-backbones",
    "topics/computer-vision-overview"
  ],
  "entities": [],
  "claims": [],
  "questions": []
}