论文阅读：DCNv2：Deformable ConvNets v2: More Deformable, Better Results

1、论文总述

这篇paper是DCNv1的升级版，文章认为，对于positive的样本来说，采样的特征应该focus在roi内，如果特征中包含了过多超出roi的内容，那么结果会受到影响和干扰。而negative样本则恰恰相反，引入一些超出roi的特征有助于帮助网络判别这个区域是背景区域。

看这篇anchor free目标检测论文 论文阅读：Objects as Points（也叫CenterNet）的时候，许多大佬的解读都是认为RepPoints是在DCNv2的进化版，因为RepPoints的卷积也都是用可变形卷积来提取特征，但训练时候对取样点有对应的损失函数，可以对DCN的自由的采样点进行监督，让其学到有用的ROI内部的特征；
还有一点就是RepPoints可以学到目标的几何特征，而DCN不能，DCN只是学到目标用来识别分类的特征，并不能用这些来进行box的回归。

从DCNv2的训练也能看出来其实，训练时候需要RCNN的mimicking来更好地让DCNV2学习目标的特征，只是分类用的。

理解这篇paper时候，发现了一个写的很好的博客，写的比较清晰而且全面，所以就不自己写了，链接为：Deformable ConvNets v2算法笔记

In this paper, we present a new version of Deformable
ConvNets, called Deformable ConvNets v2 (DCNv2), with
enhanced modeling power for learning deformable convolutions. This increase in modeling capability comes in two
complementary forms.
The first is the expanded use of deformable convolution layers within the network. Equipping more convolutional layers with offset learning capacity allows DCNv2 to control sampling over a broader range of
feature levels.
The second is a modulation mechanism in
the deformable convolution modules, where each sample
not only undergoes a learned offset, but is also modulated
by a learned feature amplitude. The network module is thus
given the ability to vary both the spatial distribution and the
relative influence of its samples.
To fully exploit the increased modeling capacity of
DCNv2, effective training is needed. Inspired by work on
knowledge distillation in neural networks [2, 22], we make
use of a teacher network for this purpose, where the teacher
provides guidance during training. We specifically utilize
R-CNN [17] as the teacher. Since it is a network trained for
classification on cropped image content, R-CNN learns features unaffected by irrelevant information outside the region
of interest. To emulate this property, DCNv2 incorporates a
feature mimicking loss into its training, which favors learning of features consistent to those of R-CNN. In this way,
DCNv2 is given a strong training signal for its enhanced
deformable sampling. （R-CNN [17] as the teacher这个我也是第一次见，对我启发很大。)

参考文献

1、Deformable Convolution v1, v2 总结

2、Deformable ConvNets v2算法笔记

3、论文阅读：Deformable ConvNets v2: More Deformable, Better Results

4、如何评价 MSRA 视觉组最新提出的 Deformable ConvNets V2？

5、目标检测论文阅读：DCN v2