1. DeepLabv2

2. PSPNet

Pyramid Scene Parsing Network

Motivation

• Mismatched Relations
• Confusion Categories
• inconspicuous classes

• Difference between theoretical and real receptive field in FCN → hard to catch contextual information

Architecture

• Average pooling applied to feature map to produce sub-regions

Global Average Pooling

• Global Average Pooling VS Convolution

  

  • Convolution outputs local context information on heat map while Global Average Pooling outputs global context information

3. DeepLab v3

• Overall architecture is similar as DeepLabv2 but for the Atrous Spatial Pyramid Pooling

• New 1x1 convolution and image pooling
• Instead of summation, DeepLabv3 concatenates heat maps

4. DeepLab v3+