Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model阅读笔记

问题：

1，representing visual data is challenging for SSMs due to the position-sensitivity of visual data 

（mamba is lack of positional awareness）

2，the requirement of global context for visual understanding

（mamba is a unidirectional model）

在MambaIRv2: Attentive State Space Restoration这篇文章中分别使用了positional embeding and "注意力机制"的引入C=(C+P),从而实现关注到全局

解决：

1，marks the image sequences with position embeddings 

这完全跟vision transformer(ViT)干的事情一样啊 

2，compresses the visual representation with bidirectional state space models.

感觉创新点的话就一个双向吧。 

本文的大框架：

ViT: