1. Semantic Segmentation

• Classify pixel
• Only care about semantic category

2. Architectures

2-1. Fully convolutional Networks

• First end to end architecture for semantic segmentation
• Arbitary input size

Fully connected vs Fully convolutional

• 1x1 convolution layer classifies every feature vector of the feature map

• Limitation : Low resolution predicted score map

  → We use upsampling

Upsampling

• Resizing small activation map to the size of the input image

• Transposed convolution
  • Limitation : Overlaps happen often

Features of FCN