LLLMWIKI
正文元数据首页

来源笔记 · 更新于 2026-07-14

Quo Vadis, Action Recognition?I3D、Kinetics 与 3D 视频预训练

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

I3D：膨胀 3D 卷积与视频预训练

一句话结论

I3D 把 2D 图像网络沿时间维扩展为 3D 卷积，并用 Kinetics 预训练建立可迁移的视频动作识别基线。

问题定义

怎样复用成熟图像网络，同时让模型直接学习时空局部模式，并克服小型视频数据集不足？

方法概述

2D 滤波器膨胀为 3D 并继承 ImageNet 权重；RGB 与光流 I3D 分别建模外观和运动，再融合。

关键发现

  • Kinetics 规模的视频预训练显著改善下游动作分类。
  • 3D 卷积可从 2D 权重初始化。
  • 强结果混合架构、图像/视频数据与光流，不能单因素外推。

局限或疑问

  • 3D 卷积成本高，clip 上下文有限。
  • 光流流仍需要外部计算。
  • 动作分类不能代表长视频理解。

原始链接

相关页面

元数据
{
  "id": "2026-07-14-i3d",
  "type": "source",
  "title": "Quo Vadis, Action Recognition?：I3D、Kinetics 与 3D 视频预训练",
  "status": "reviewed",
  "created": "2026-07-14",
  "updated": "2026-07-14",
  "venue": "CVPR 2017",
  "ingested_at": "2026-07-14",
  "tags": [
    "near-cvpr-2025",
    "video-representation",
    "video-understanding",
    "representation-learning",
    "primary-source"
  ],
  "note_status": "reviewed",
  "source_type": "paper",
  "authors": [
    "Joao Carreira",
    "Andrew Zisserman"
  ],
  "published_at": "2017-01-01",
  "canonical_links": [
    "https://arxiv.org/abs/1705.07750",
    "https://openaccess.thecvf.com/content_cvpr_2017/html/Carreira_Quo_Vadis_Action_CVPR_2017_paper.html",
    "https://arxiv.org/pdf/1705.07750",
    "https://github.com/deepmind/kinetics-i3d"
  ],
  "raw_entry": "raw/ingest/2026-07-14-i3d/",
  "analysis_note": "raw/ingest/2026-07-14-i3d/analysis.md",
  "topics": [
    "topics/video-representation-and-temporal-modeling",
    "topics/video-understanding"
  ],
  "entities": [],
  "claims": [],
  "questions": []
}