R-FCN网络学习
由于传统的检测网络:apply a costly per-region subnetwork hundreds of times
R-FCN:almost all computation shared on the entire image.
解决问题:分类的平移不变性和检测的平移敏感性。(a dilemma between translation-invariance in image classification and
translation-variance in object detection)
介绍:
流行的检测网络一般是由RoI分割成两个子网络:(1)独立于RoIs的全卷积子网络(共享运算)
(2)RoI-wise子网络(不共享运算)
历史问题:AlexNet and VGG Nets 这种传统网络的spatial pooling layer is naturally turned into the RoI pooling layer在目标检测中。
最近的新网络:ResNets and GoogLeNets are by design fully convolutional. 在目标检测架构中,很自然的使用全部的卷积层来构建共享卷积子网络,让RoI-wise没有隐藏层。但是:降低了检测精度,不能与分类精度匹配。为了解决这个问题:ResNet paper里面提到Faster R-CNN的RoI pooling layer is unnaturally inserted between two sets of convolutional layers,breaks down translation invariance(the post-RoI convolutional layers are no longer translation-invariant when evaluated across different regions),提高了精度,但是由于unshared per-RoI computation牺牲了速度。since it introduces a considerable number of region-wise layers
each of the kxk bin aggregates responses from only one score map out of the bank of kxk score maps 。
要实现平移敏感,引入了position-sensitive score maps,通过使用一堆 specialized convolutional layers。