MetaFormer is Actually What You Need for Vision 

In this video, we explore the revolutionary MetaFormer architecture, which extends the highly successful transformer idea from natural language processing into the world of computer vision. MetaFormer introduces a general architecture block that guarantees exceptional performance, regardless of the specific task at hand.
To demonstrate this claim, the researchers behind MetaFormer present an embarrassingly simple yet incredibly effective architecture called PoolFormer. Surprisingly, this architecture outperforms the popular Swin Transformer model with significantly fewer parameters and computations.
Join us to learn about the exciting potential of MetaFormer in transforming the field of computer vision and enabling breakthroughs in a wide range of applications.
Paper link: arxiv.org/abs/2111.11418
Table of Content:
00:00 Introduction
01:46 ConvNeXt resemblance
02:48 PoolFormer block
06:10 Model Architecture
08:16 Result
Icon made by Freepik from flaticon.com