LLLMWIKI
ArticleMetadataMain page

Source note · Updated 2026-07-14

Two-Stream Convolutional Networks for Action RecognitionRGB 与光流双流视频表示

#near-cvpr-2025#video-representation#video-understanding#representation-learning#primary-source

Two-Stream ConvNets：外观与运动双流

一句话结论

Two-Stream 用 RGB 空间流与光流时间流分别建模外观和运动，建立视频动作识别的经典双流表示。

问题定义

单帧外观不足以区分许多动作时，怎样把短时运动明确提供给卷积网络？

方法概述

空间流读取 RGB，时间流读取连续稠密光流堆叠，两者独立训练并在分类分数层融合。

关键发现

  • 显式运动表示补充单帧外观。
  • 图像预训练帮助空间流，小视频数据限制时间流。
  • 双流分解影响后续视频模型设计。

局限或疑问

  • 预计算光流昂贵且会传递误差。
  • 晚期融合缺少细粒度跨流交互。
  • 短片分类不等于长视频时序推理。

原始链接

相关页面

Metadata
{
  "id": "2026-07-14-two-stream-convnet",
  "type": "source",
  "title": "Two-Stream Convolutional Networks for Action Recognition：RGB 与光流双流视频表示",
  "status": "reviewed",
  "created": "2026-07-14",
  "updated": "2026-07-14",
  "venue": "NeurIPS 2014",
  "ingested_at": "2026-07-14",
  "tags": [
    "near-cvpr-2025",
    "video-representation",
    "video-understanding",
    "representation-learning",
    "primary-source"
  ],
  "note_status": "reviewed",
  "source_type": "paper",
  "authors": [
    "Karen Simonyan",
    "Andrew Zisserman"
  ],
  "published_at": "2014-01-01",
  "canonical_links": [
    "https://arxiv.org/abs/1406.2199",
    "https://papers.nips.cc/paper_files/paper/2014/hash/00ec53c4682d36f5c4359f4ae7bd7ba1-Abstract.html",
    "https://arxiv.org/pdf/1406.2199"
  ],
  "raw_entry": "raw/ingest/2026-07-14-two-stream-convnet/",
  "analysis_note": "raw/ingest/2026-07-14-two-stream-convnet/analysis.md",
  "topics": [
    "topics/video-representation-and-temporal-modeling",
    "topics/video-understanding"
  ],
  "entities": [],
  "claims": [],
  "questions": []
}