Lecture 2 : Image Classification Pipeline

• Computer Vision이 어려운 이유
  1) Viewpoint variation
  2) Illumination
  3) Deformation
  4) Occlusion - 가려져있는 것/일부만 보이는 경우
  5) Background clutter
  6) Intraclass variation
  
  
• Rule-based vs Data-Driven Approach
  - Rule-based - Hand crafted features
  1) edge&corners : 픽셀값이 갑자기 변하는 경우(미분해서)
  2) Bag of words
  3) HOG(Histogram of Oriented Gradient for human detecting)
  
  - Data-Driven
  Deel Learning  Alg. === [Feature Extractor -> Features + ML Alg.]
  
  
• KNN (K-Nearest Neighbors)
  - Distance Metric: L1 distance & L2 distance
  

  

  
  - very slow at test time
  - curse of dimensionality 차원의 저주 --> 일부만 뽑아서.. Manifold learning...
  - distance metrics on pixels are not informative
  - KNN on images never used!
  
  
• Data Split
  [train+validation]+test
  

  

• Linear Classification