CODEWITHSUNDEEP
machine-learningcomputer-visiondeep-learning
user288609
user288609

Reputation: 13035

the relationship between a specific layer and its receptive field

In one paper, I saw the following table regarding the relationship between layers and their respective receptive fields. I have two questions:

1) How to calculate this kind of relationship, are there nay formulas to follow.

2) For instance, in the last several layers, such as pool5, fc6&fc7, have pretty large receptive fields. The receptive field for fc6&fc7 is 404*404. If the training image is 256*256, does it mean that fc6&fc7 see a lot of pixels out of the boundary of 256*256 ( or just waste the power of fc6&fc7 which has 404*404 as receptive field). I am not sure how to understand the physical meaning of receptive field in terms of a specific layer when performing classifciaiton/feature extraction type of work.

enter image description here

Upvotes: 1

Views: 108

Answers (1)

Amitay Nachmani
Amitay Nachmani

Reputation: 3279

Lets say you have 2 layers with max pulling in between. The first layer will be a convlutional layer with filter 3X3 and a stride of 1X1 than max pooliing layer of 2X2 wit a 2X2 stride and than a second convlutional layer with a filter size of 3X3 and stride of 1X1.

The first layer looks at neiogbhords of 3X3 from the oriignal image so the recptive field will be 3X3. Than the max pooling layer effectvilly downsample the image by a factor of two so the input to the second layer is the origianl image dowsampled by a factor of 2. Now the current layer also look for at neigbhords of 3X3 but ON THE DOWNSAMPLED image which means that effectavilly this filter looks at pixels that came from a (downsapleFactorfilterSize)X(downsapleFactorfilterSize) niegbrhood of the original image!!! so the recptive field will be 6X6.

Upvotes: 0

Related Questions