1. Motivation

• There are three challenges in semantic segmantation with DCNN. (1) reduced feature resolution, (2) existence of objects at multiple scales, and (3) reduced localization accuracy due to DCNN invariance.
• The first challenge is caused by max-pooling and down sampling. To address this problem, the paper removed the downsampling operator and upsample the filters. In the other words, atrous convolution is used for recovering feature map.
• To deal with second challenge, the model used astrous spatial pyramid pooling
• The last challenge is related with the fact that traditioanal classifiers required transformation invariance. This paper used CRF to improve the model's ability to capture fine features.

2. Method

2.1. Atrous Convolution for Dense feature Extraction and Field-of-View Enlargement

• The paper mentioned that using astrous DCNN is more effective for feature resolution
  

2.2. Multiscale Image Representations using ASPP

• As following Fig 4, the paper have implemented a model which use multiple parallel atrous convolutional layers.
  

2.3. Structured Prediction with Fully-Connected Conditional Random Fields for Accurate Boundary Recovery

• To recover local fine boundary, the model used short-range CRF