Try   HackMD

Object Detection Develop History-1

Two Stage Method
stage 1 = region proposal
stage 2 = feature extract + bounding box

R-CNN
Fast R-CNN
Faster R-CNN
R-FCN


R-CNN
用傳統的computer vison 的selective search找region proposal
但是stage 2用CNN找FEATURE + BOUNDING BOX + SVM辨識目標種類

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

缺點: 太慢


Fast R-CNN

改進: 每個region proposal共用CNN來找feature&SVM換成Neural Network來加速

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

缺點: 後面分類物體 + 給bounding box變快,前面的region proposal就顯得很慢


Faster R-CNN

改進: region proposal也用CNN來加速

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

缺點: 每個proposal區域的分類還是要單獨算一次


R-FCN

改進: 整張影像的分類只要算一次

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →


Two Stage Method Pros and Cons

Accuracy: High
Speed: Slow(Relatively)

因此產生了不再分兩個階段的method -> One Stage Method

One Stage Method

Accuracy: Low
Speed: fast(Relatively)
ex. YOLO
將兩個階段的CNN合併成一個網路,同時選取proposal + classification

但是,還有一種加速two stage的method-> Light weighted

Light weighted Two Stage

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

ex. PVANet


Summary

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

YOLO已經有5代了

tags: SAR Deep learning Object Detection review