Firstly as an example here is the architecture of YOLOv2 enter image description here

I am trying to understand the depth of an output of a convolutional layer. For example, the first convolutional layer has the shape 3x3x32. So there are 32 filters with shape 3x3, but each filter has 3 layers and these 3 layers convolve over 3 layers of the input. At the end, values of the 3 layers are summed up and to generate 1 layer. For 32 filters, we get an output with 32 layers.

If we look at the next layer, 64 filters with size 3x3 and each filter should have 32 layers. Because input has 32 layers. Is this inference true? If it is not, how does it work?

  • 1
    $\begingroup$ I think that this question is a duplicate of this. Please, let me know whether that is the case or not. In any case, to answer your question: yes, your reasoning is correct, but note that this only applies to 2d convolutions (the usual convolution operation in CNNs). In the case of 3d convolutions, as I write in my answer in the linked post, the depth of the kernels could be different than the depth of the inputs. $\endgroup$
    – nbro
    Dec 28, 2020 at 15:43
  • 1
    $\begingroup$ @nbro basically yes, they are same. Sorry, I did not see that one before asking mine. $\endgroup$
    – Reactionic
    Dec 28, 2020 at 15:49