引文范围
Per-pixel动态卷积:[3, 20, 22, 37, 45, 49, 59]
具体内容
引文3
一句话总结: 可微分的区域分组per-pixel动态卷积
实验结论: 1. 小模型收益更高;2. 区域越多收益越高
Title: [CVPR 2021] Dynamic Region-Aware Convolution
Paper: https://arxiv.org/abs/2003.12243
Code: <unofficial!!> | https://github.com/shallowtoil/DRConv-PyTorch
整理:[知乎] | [CVPR 2021] Dynamic Region-Aware Convolution | https://zhuanlan.zhihu.com/p/372120226
引文20
一句话总结:
Title: [arXiv 2021.7] Laconv: Local adaptive convolution for image fusion用于图像融合的局部自适应卷积
Paper: https://arxiv.org/abs/2107.11617v1
Code: None
整理:None
引文22
Title: [CVPR 2021] Involution: Inverting the inherence of convolution for visual recognition
Paper: https://arxiv.org/pdf/2103.06255.pdf
Code: https://github.com/d-li14/involution
整理:[知乎] | CVPR 2021 | 真内卷!Involution:反转卷积的固有性以进行视觉识别 | https://zhuanlan.zhihu.com/p/356369472
引文37
Title: [CVPR 2019] Pixel-adaptive convolutional neural networks
Paper: https://arxiv.org/abs/1904.05373
Code: https://suhangpro.github.io/pac/
整理:[CSDN] | 每天一篇论文 308/365 Pixel-Adaptive Convolutional Neural Networks | https://blog.csdn.net/qq_26623879/article/details/103849325
引文45
Title: [ICCV2019] Carafe: Content-aware reassembly of features
Paper: https://arxiv.org/abs/1905.02188
Code: mmdetection | https://github.com/open-mmlab/mmdetection
整理:[CSDN] | 目标检测系列:CARAFE: Content-Aware ReAssembly of FEatures
引文49
Title: Adaptive convolutions with per-pixel dynamic filter atom
Paper: https://arxiv.org/abs/2108.07895
Code: None
整理:None
引文59
一句话总结: 空间和通道维度分离生成核权重的动态卷积
Title: [CVPR 2021] Decoupled dynamic filter networks
Paper: https://thefoxofsky.github.io/files/ddf.pdf
Code: https://github.com/thefoxofsky/ddfnet
整理:[知乎] | 【CVPR2021】Decoupled dynamic filter networks | https://zhuanlan.zhihu.com/p/459147690
总结
这些动态卷积的论文,总结下来方法上有相似的操作:
- GAP很常用,用来浓缩沿某一个维度的特征;
- GAP之后经常接着权重生成网络,基本是Conv+BN+ReLU+Act+Conv的格式,其实大部分用FC?ReLU和Sigmoid都看到有用。
- 经常有矩阵分解(我理解和深度可分离、空间可分离的动机、形式都差不多)来尽可能地做轻量化;
结论上,有以下相似之处:
- 基本是在小模型上,替换深层次模块更有利——这其实挺符合假设的,大模型通过blindly地增加channel数,很可能也已经把有效的kenerl模式都蒙出来了