Two Stream

Two-Stream Convolutional Network for Action Recognition in Videos

Inspiration for me

Problems raised in this article, research directions to explore, new ideology

• Video provides a natural wiki.en data augmentation
• 双向，金字塔结构或cascade

Basic Thought

• Two type of information
  • appearance
  • motion between frames
• Video is a type of natural wiki.en data augmentation
• 人眼
  • ventral stream → object recognition
  • dorsal stream → recognize motion
• optical flow → dense trajectory. L帧的视频得到L-1帧的光流图。通过相邻帧分析，每个点都会有一个运动矢量。
• 在本文之前的网络都没有学到时序信息。
• How to improve ?
  • Late Fusion -> Early Fusion ?
  • AlexNet -> VGG, NiN, wiki.en LSTM...
  • Single Frame -> long video processing
  • Optical flow
    • sample along the trajectory. result shows it's worse
    • Bi-direction
    • sparse光流，体积极大变小光流

model architecture

new points

What's new raised by this article ?

• New model
  • Why？Where ?New architecture ? How？Important
• New ideology
• New function
• New mission
• New training method

advancement

The progress gained based on article

• Beyond Short snippets: Deep Networks for Video Classification
  • Google
  • wiki.en LSTM → in short snippets, not so effective as we may thought
  • optical flow