阅读笔记-PVT-Pyramid Vision Transformer_A versatile backbone for dense prediction without convolutions
来源:arXiv:2102.12122v1单位:南大、南理、商汤、港中文代码:https://github.com/whai362/PVTtitle文章内容用一句话概括就是给ViT方法装上金字塔结构处理密集预测问题。主要创新点包括两点:1.progressiveshrinkingstrategy能够实现金字塔结构;2.spatialreductionattention减少self-attentio