1. Introduction

• This paper proposed co-FCN network which make conditioning branch contain few-shot annotations. Samples for few-shot learning is selected from segmentation datset.
  

2. Conditional Architecture and Few-Shot Optimization

• Few-shot segmentation is required to learn to segment new input with a few annotations.
• In this paper, proposed model based on FCN and it align few shot trainng and testing paradigms.
• Figure 1 shows overall structure of co-FCN. The segmentation branch is a FCN that segments the query. The conditioning branch takes support sets as input and makes features which are fused with query image.

4. Discussion

• The proposed model is much faster then tranditional model, because it doesn't need any optimization at test time. Additionally, it is good to cope with sparse annotations.

Reference

• CONDITIONAL NETWORKS FOR FEW-SHOT SEMANTIC SEGMENTATION