Padding for conv filters

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Starting with the idea of using padding to increase the contributions of the border pixels to the filtered image, the course talks about padding with 0. It seems that this would create a distinct border artifact around the whole filtered image. There are other ways to pad, such as reflecting to values of border pixels in the original into the padding positions.

It seems that there will always be a border artifact – there are only ways to reduce its severity.

My questions:

  1. Is this artifact of great significance to the downstream algorithm processing blocks?

  2. For applications used in processing road scenes as part of automated cars, are other methods of padding indicated – such as the reflection method above or others?

Thank you!
–Paul

2

I think the main purpose of padding is so that the feature set doesn’t shrink in the convolution layers.
The pad values don’t matter a great deal, because they are constants for all of the pixels.

2 Likes