Boostcamp AI Tech (Day 032)

My assignment: Semantic Segmentation

CNN architectures (2)

• Deeper layers

  • Larger receptive fields
  • More capacity and non-linearity
  • Gradient vanishing/exploding
  • Computationally complex
  • Degradation problem

• GoogLeNet (CVPR 2015)

  • Inception module
    • 1x1 Conv
    • 1x1 Conv $\rightarrow$ 3x3 Conv
    • 1x1 Conv $\rightarrow$ 5x5 Conv
    • 3x3 pool $\rightarrow$ 1x1 Conv
    • Concatenate all filter (channel axis)
  • 1x1 Conv (bottleneck)
    • reduce the number of channels
  • Architecture
    • Stem network (vanilla CNN)
    • Stacked inception modules
    • Auxiliary classifiers
      • injecting additional gradients into lower layers
      • used only during training

• ResNet (CVPR 2016)

  • 152 layers
  • Degradation problem (optimazation problem)
    • Solve with Residual block
  • Residual block
    • Shortcut connection (skip connection) ($+ x$)
  • Architecture
    • He initialization (not too high)
    • 7x7 Conv at the beginning
    • Stack residual blocks
      • 3x3 Conv + Batch Norm
    • Block 마다 한 번씩 stride 2
      • doubling no. of filters
      • spatially down-sampling
    • Single FC layer

• Beyond ResNet

  • DenseNet
    • 훨씬 이전의 layer의 정보도 skip connection
    • concate along channel axis
  • SENet (Squeeze and Excitation)
    • Recalibrate channel-wise reponses by modeling interdepenencies b/w channels
    • SE
      • Squeeze: capturing distributions of channel-wise reponses by global average pooling
      • Excitation: gating channeling by channel-wise attention weights obtained by a FC layer
  • EfficientNet
    • Compound scaling
      • width scaling
        • more channels (GoogLeNet, DenseNet)
      • depth scaling
        • deeper layers (ResNet)
      • resolution scaling
        • higher input resolution
  • Deformable convolution
    • standard CNN + grid sampling with 2D offsets
      
      

Semantic segmentation

• Definition

  • Classify each pixel of an image into a semantic category
  • Applications
    • Medical image
    • Autonomous driving
    • Computational photography

• FCN (CVPR 2015)

  • Fully Convolutional Networks
  • first end-to-end architecture for semantic segmentation
  • Input/Ouput
    • Input: arbitrary size
    • Output: segmentation map (corresponding size to Input)
  • Difference from Fully connected
    • output has spatial coordinates (heat map)
  • Upsampling
    • 일단 작게 만들어 receptive field 키운 뒤 upsampling을 통해 input과 같은 resolution으로 만들어줌
    • Transposed convolution만 수행시 overlap issue 발생
      • Nearest-neighbor, Bilinear와 같은 interpolation을 추가적으로 수행
  • Add skip connections for enlarging score map
    • Layers 차이
      • lower
        • fine, low-level, detail, local
      • higher
        • coarse, semantic, holistic, global
    • Integrates activations from lower layers into prediction
    • Preserves higher spatial resolution
    • Captures lower-level semantics at the same time

• U-Net (MICCAI 2015)

  • 특징
    • Share same FCN property
    • Predict a dense map by concat feature maps from contracting path
      • similar to skip connections in FCN
    • Yield more precise segmentations
  • Architecture
    • Contracting path
    • Expanding path
  • Skip connection
    • Concatenation of feature maps provides localized information

• DeepLab (ICLR 2015)

  • 참고 자료
  • CRFs (Conditional Random Fields)
    • 1st row: score map (before softmax)
    • 2nd row: belief map (after softmax)
  • Dilated(확장된) convolution
    • Atrous convolution
    • Insert spaces b\w kernel element (dilation factor)
    • Exponentially expand receptive field (same parameter number)
  • Depthwise separable convolution
    1. Depthwise convolution
    2. Pointwise convolution