1. R-CNN

1. Input image
2. Extract Regional Proposals
   • Selective Search
   • Sliding Window
3. Compute CNN features
4. Classify Regions

1-1. Architecture

• Bbox regression → selective search를 이용하여 구한 ROI를 ground truth에 맞추는 과정

1-2. Training

• AlexNet
• Linear SVM
  • Hard negative mining
• Bbox regressor

1-3. Shortcomings

• 2000개의 region을 각각 CNN 통과
• 강제 warpping, 성능 하락
• CNN, SVM classifier, Bbox regressor 따로 학습
• End-to-End X

2. SPPNet