Topic: Proposal by Region Guided Anchoring
arXiv: https://arxiv.org/abs/1901.03278
Reference article: https://zhuanlan.zhihu.com/p/55854246
basic introduction
This paper mainly to improve on the shortcomings of existing RPN, he said improvement is not accurate, because the two methods are very different, but when they do the same job, is to get the candidate region. .
This paper proposed a new method does not require the use of pre-set anchor box
This paper introduces
RPN shortcomings
(1) anchor the required dimensions and aspect ratio of pre-defined, which is a relatively large impact on performance parameters over, and need to be adjusted individually for different data sets and methods. If the scale and aspect ratio setting is not appropriate, may cause the recall is not high enough, or anchor the excessive influence of the classification performance and speed;
(2) most of the anchor are located in the background region, will not have any positive effect on the proposal or detection ;
(3) predefined anchor shape may not be able to meet the extreme disparity in size or the aspect ratio of the object.
GA-RPN
This is the overall structure of the model, the main focus is on the right side guided anchoring this section, which is the essence of this paper.
Central point forecast
在拿到feature map 之后，传统方法就开始进行滑动窗口，把每个像素点都当做中心产生候选区域，前边已经说过这种方法的缺点，大量背景也产生了大量无效的候选区，如何避免这个问题呢？
我们需要有针对行的把一些像素点当做中心就可以了，具体怎么做呢，就是上图的NL部分，通过对feature map进行1×1的卷积，然后对卷积后的值进行sigmoid，就得到了一个概率图（probability map），而概率图中每个点的值代表这个点为目标中心点的概率，然后设定一个阈值，大于某个概率即说明这个点很有可能作为某个目标的中心点，接下来以这个点为中心，针对性的产生候选框即可，从而避免了滑动窗口的缺点。
那么如何学习（训练）到这样一个模型呢？
论文将整个 feature map 的区域分为物体中心区域，外围区域和忽略区域，大概思路就是将 ground truth 框的中心一小块对应在 feature map 上的区域标为物体中心区域，在训练的时候作为正样本（target = 1），其余区域按照离中心的距离标为忽略或者负样本（target = 0），这样，有了目标值，训练即可，论文使用Focal loss来训练这个分支。
这样我们就得到了一个产生中心点概率图的模型NL，接下来要在中心点上产生边框（确定w和h的值）
形状预测
形状预测模型为上图中的Ns部分，具体是采用1×1的卷积核，产生一个双通道的feature map，但直接预测w和h不好预测，你所以转而预测dw和dh，之后经过公式转换得到w和h，公式如下：
拿如何训练这个预测模型呢？
训练过程最重要的是要找到w和h的目标值，对于以前常规的 anchor，w和h是事先规定好的，我们可以直接计算它和所有 ground truth 的 IoU，然后将它分配给 IoU 最大的那个 gt，gt的w和h即为目标值。但是很不幸现在的 anchor 的 w 和 h 是不确定的，是一个需要预测的变量，该怎么办呢？sample anchor。
论文将这个 anchor 和某个 gt 的 IoU 表示为：
我们不可能真的把所有可能的 w 和 h 遍历一遍然后求 IoU 的最大值，所以采用了近似的方法，也就是 sample 一些可能的 w 和 h。理论上 sample 得越多，近似效果越好，但出于效率的考虑，所以 sample 了常见的 9 组 w 和 h。
形状预测损失函数：
这样Ns位置预测分支也准备好了可以画框了。
可以看到，anchor 基本都集中在有目标的区域，确实产生了不错的效果。
Feature Adaption 模块
但是我们发现一个不合理的地方，大家都是同一层 conv 的特征，凭啥我就可以比别人优秀一些，代表一个又长又大的 anchor，你就只能代表一个小小的 anchor。
The reason is that the unreasonable one hand, in a different location in the same layer of conv, feature of receiptive field is the same as in the original RPN inside, we have said anchor the same shape, so no problem, but now each has its anchor its own unique shape and size, and the feature is not particularly good match. On the other hand, the characteristics of the original map, it does not know the shape of the prediction branch prediction anchor shape, but the next classification and regression is based on the predicted anchor to do, it may be more ignorant force.
We've added a Feature Adaption module to solve this problem, the anchor shape information integration directly into the feature map, a new feature maps so you can get to adapt to the shape of each anchor position. We use a 3x3 deformable convolution to FIG corrected original feature, the deformable convolution is offset by the anchor w and h obtained through a 1x1 conv. (Where the focus should be designated, if it is the same as a normal deformable convolution, characterized by predicting FIG offset, the increase is limited, since there is no play according to adapt to the shape of the anchor effect)
By so doing, to allow the effective range of the feature shapes and anchor closer to the object, a different location in the same conv anchor may be representative of the various shapes and sizes.