Day32 학습정리 - CV2

강의복습

1. Image classification II

1) Problems with deeper layers

• Gradient vanishing / exploding
• Computationally coomplex
• Degradation

2) CNN architectures for image classification 2

• GoogLeNet
  • 채널 개수 줄이기 위해 1×1 convolution 적용
  • Stem network
  • Stacked inception modules
  • Auxiliary classifiers
  • Classifier output (a single FC layer)
• ResNet
  • Degradation 문제 해결하기 위해 Shortcut connection 적용
• 기타
  • DenseNet
  • SENet
  • EfficientNet
  • Deformable convolution

3) Summary of image classification

• AlexNet
  • Simple CNN architecture
  • Simple computation, but heavy memory size
  • Low accuracy
• VGGNet
  • simple with 3×3 convolutions
  • Highest memory, the heaviest computation
• GoogLeNet
  • inception module and auxiliary classifier
• ResNet
  • deeper layers with residual blocks
  • Moderate efficiency(depending on the model)
• Backbone model: 주로 VGGNet 또는 ResNet이 사용됨

2. Semantic segmentation

 1) Semantic segmentation

• 이미지의 각 픽셀을 카테고리로 분류

2) Semantic segmentation architectures

• Fully Convolutional Networks(FCN)
  • first end-to-end architecture for semantic segmentation
  • Fully connected layer: fixed dimensional vector 출력, 위치정보 버림 (Image classification)
    Fully convolutional layer: classification map 출력 (Semantic segmentation)
  • 1×1 convolution layer의 한계: 해상도 매우 낮음
    →upsampling으로 해결
  • Upsampling: 작아진 이미지를 input image 크기로 복원 (참고: ronjian.github.io/blog/2018/03/23/CNN)

Transposed convolution	Upsample and convolution
Checkerboard artifacts	overlap issue 피함 Nearest-neighbor(NN) / Bilinear interpolation → convolution

• skip connection (참고: modulabs-biomedical.github.io/FCN)

• Hypercolumns for object segmentation
  • Hypercolumn: stacked vector of all CNN units on that pixel
  • FCN과 유사한 구조 (차이점: 각 bounding box마다 적용)

• U-Net
  • Contracting Path: 3×3 convolution 계속 적용, feature channel × 2, feature map 크기 × 1/2
    Expanding Path: 2×2 convolution 계속 적용, feature channel × 1/2, feature map 크기 × 2
  • input, feature size 짝수여야 함

• DeepLab
  • Conditional Random Fields (CRFs)
    • 1st row: score map (before softmax)
    • 2nd row: belief map (after softmax)

---

코멘트

오늘 피어세션에서 팀원분들이 나는 미처 생각 못했던 좋은 질문들을 많이 하셔서 나도 앞으로 공부할 때 대충 넘기지 말고 더 꼼꼼하게 해야겠다고 생각했다.