Why do advanced convolutions preserve resolution?

Question

Animation from here . I wonder why the advanced convolution maintains resolution, it is claimed. The blue input appears to be 7x7 and green is 3x3.

EDIT:

One way to get around the loss of resolution is to insert an input with approximately half the size of the current susceptible field, but

• this substantially undermines the claim that extended convolutions do not lose permission because it is an add-on that preserves permission. To get the same output size with the input, a regular convolution requires even less filling.
• , - . , 1024x1024 10- , 2048x2048 (, , ). 4- , , . .

Answer 1

5x5. 3x3, 0, .

, , :

. .

. , , 3 3x , 15x15 .

3x3 , 3 (4 ) ( , , , ). 4 3x3 3 3, 15x15.

, , , .