1. Spatial Transformer Layer

1.1 CNN is not invariant to scaling and rotation

(1) CNN并不能真正做到scaling和rotation.
(2) 如下图所示,在通常情况下,左右两边的图片对于CNN来说是不一样的.

  所以,我们考虑一层layer,这层layer能够对input image进行旋转缩放,以便更好地识别.
  如上图所示,Spatial Transformer Layer是Neuron Network,而它的作用是多学习一层layer,对左边的图片做scaling和rotation后,能够被CNN识别出来.当然,这个Layer也可以transform CNN的feature map.

1.2 How to transform an image/feature map

本栏目推荐文章
• LLM series: Transformer
• Learning Dynamic Query Combinations for Transformer-based Object** Detection and Segmentation论文阅读笔记
• [NLP复习笔记] Transformer
• RNN 和 Transformer 复杂度比较
• 挑战Transformer的新架构Mamba解析以及Pytorch复现
• 《Span-Based Joint Entity and Relation Extraction with Transformer Pre-Training》阅读笔记
• ICLR 2022: Anomaly Transformer论文阅读笔记+代码复现
• AMOS: Enabling Automatic Mapping for Tensor Computations On Spatial Accelerators with Hardware Abstraction
• Visual Transformer 与归纳偏置
• Swin Transformer

Transformer Spatial 3.3transformer spatial 3.3 spatial spatial-temporal 3.3 cvpr_spatial-frequency reference spatial change export 练习题3.3 线性3.3 spatial空间join 代码3.3