1. Motivation

• Semi- and weakly supervised segmentation methods cannot segment for a new input class. Therefore, this paper attempted to perform segmentation of a new input given a few support image in the same class.
• To perform such problem, this paper addressed three key parts of FSS problems (1) how to summarize the sparse, structured support into a task representation , (2) how to condition pixelwise inference on the given task representation, and (3) how to synthesize segmentation tasks for accuracy and generality

2. Method

• The work deal with the problem with the model with two branches, (1) a guide branch for extracting the task representation from the support and (2) an inference branch for segmenting queries given the guidance.
  

2.1. Guidance: From Support to Task Representation

• There is a problem in early fusion: In compatibility of the support and query representations. The paper address is problem like Figure 3 (b).
• In late fusion, first it extract feature from the encoder, map the annotations into masks in the same channel with the feature map. Finally fuse feature map by multiplication.

2.2. Guiding Inference

• 

3. Result