Hello everyone I have the following architecture from the DELP-DAR research paper (https://www.sciencedirect.com/science/article/pii/S0167865519303216)

network architecture

and I dont really understand two things, first how can you go from 960x570 size to 481x286 with a 3x3 convolution? Does this mean I have to have asymmetric padding or what?

Second, what is the value after a slash? 3 x 3 / 2 - i though it to be stride but then you cant have a stride of 2 and end up with output of the same dimensions as the input.

Thank you for your help.


Ok, so I think I figured it out with the help of some micro-chart I didn't notice earlier. conv 1x1/1 is a layer with the input from the prev layer( lets call it x), the conv 3x3/2 is another layer with the input from the same 'x', while the conv 3x3/1 is a layer stacked on the conv 1x1/1, the outputs from 3x3/1 and 3x3/2 are then concatenated.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.