1. Introduction

• 이전에 convnet을 이용한 semantic segmentation은 각 pixel을 둘러싸는 물체나 지역으로 구분되지만, 이러한 과정에서 분명한 단점 존재 (낮은 성능, 오래 걸림 등)

  • Patchwise training : 효율적이지 않음

• Fully Convolution Network (FCN)

  • end-to-end, pixels-to-pixels 학습
  • pixelwise prediction, supervised pre-training
  • input size에 구애받지 않음
  • 전체 이미지가 동시에 수행됨
  • VGG, Inception with pre-trained for ImageNet 사용
  • "skip" architecture

  

2. Related work

Fully convolutional networks

• Sliding window detection
• Semantic segmentation
• Image restoration

Dense prediction with convnets

• Small models restricting capacity and receptive fields
• Patchwise training
• Post-processing by superpixel projection, random field regularization, filtering, or local classification
• Input shifting and output interlacing for dense output as introduced by OverFeat
• Multi-scale pyramid processing
• Saturating $\text{tanh}$ nonlinearities
• Ensembles

3. Fully convolutional networks

• Fully convolutional network (FCN)
  • nonlinear function을 계산하는 기존 방법들과 달리, nonlinear filter를 계산
  • loss function은 spatial dimensions의 합 ($l(\mathbf{x};\theta)=\sum_{ij}l'(\mathbf{x}_{ij};\theta)$) → gradient도 각 spatial component의 gradient의 합

3.1. Adapting classifiers for dense prediction