1. Motivation

• This paper tried to upgrade accuracy of scene segmentation by viewing the image regarding context prior.
• The problem of model based on FCN is lack of catching global scene category clues.
  
• To deal with such problem, this paper proposed PSPNet for pixel prediction

2. Method

2.1 Pyramid Pooling Module

• This module will help the model to catch global context
• Global averaging pooling may be good baseline model for global contextual pripor. However it is not enough for scene segmentation.
• The paper proposed pyramid pooling module for global scene

• As you can see in feature 3, the model reduced pyramid layer to 1 and cancatenated as the final pyramid pooling global feature.

2.2. Network Architecture

• First, the model extract feacture model with Resnet. After then, it use pyramid pooling module to gather context information. Then, the model concatenate the prior with the original feature map.

3. Results