【CNN】|How much position information do convolutional neural networks encode?

ICLR2020

CNN Whether the extracted features contain location information

CNN The location information is encoded implicitly , And with the increase of network layers and convolution kernel , That is, the increase of receptive field , It can better encode location information . among , This location information is provided by zero-padding Caused by the , Image edge zero-padding Provides the boundary information of the image . Originally , The network does not know the position of each pixel or feature point . however , adopt padding Of zero, Provide a relative position information to the model , Know the distance of each feature point zero Distance information of the boundary .

There is zero-padding Under the circumstances , be based on VGG and ResNet All the models can predict the reasonable position related output , Such as abscissa or ordinate .

In the absence of padding Under the circumstances , The output will only respond directly to the input , You cannot predict location information that is not related to content .

The most direct way is, of course, to set the coordinates of each pixel concat To input or intermediate features , This simple and direct approach can be found in SOLO[3] The result of instance segmentation 3.6 AP The promotion of .