Aggregating Features

• Feature concatenation
  • Channel-dimension concatenation
  • Spatial dimension concatenation
  • Size increased
  • Example
    • GoogleNet (128 + 192 + 96 + 256 => Filter Concatenation => 672
    • Spatial Pyramid P Net (16 + 4 + 1)
    • U-Net
• merging (elementwise)
  • Size remains same
  • Example
    • Feature Pyramid Network
      • Merged with the corresponding bottom-up map (which undergoes a 1x1 convolutional layer to reduce channel dimensions)
      • 이미지를 다양한 scale로 조절하여 다양한 feature를 추출, 크기에 따라 상대적으로 이용하여 predict 진행
      • Single-scale image input으로부터 multiple level에서 resizing된 feature map을 출력할 수 있는 architecture.
      • Bottom-up pathway(CNN의 feed-forward 방식 computation) - CNN의 각 stage
      • ResNet : input dimension 유지