ChengYue's Cat

Simpler is Better: Few-shot Semantic Segmentation with ClassifierWeight Transformer

Simpler is Better: Few-shot Semantic Segmentation with ClassifierWeight Transformer 2021_ICCV

论文简要介绍

模型概览

模型包括两个训练阶段:

作者说这样做的动机是 The intuition is that, the pairs involving a query image pixel from the new class often enjoy higher similarity than those with background classes except few outlier instances; as a result, this attentive learning would reinforce this desired proximity and adjust the classifier weights conditioned on the query. Consequently, the intra-class variation can be mitigated.

实验结果

消融实验