In video, it said that when if the filtering matrix is contained in the image of image plus padding do calculation else do not perform calculation. If not filtering matrix is not contained in image or image plus padding then how to compute data on output image for specific pixel?
The whole point of convolution is to condense image information in a smaller matrix, if the image is smaller than the filter that it defeats this purpose, then there is no need to perform the convolution!
3 Likes
Right, at Gent says you can only compute outputs when the filter is completely contained within the input image. It wouldn’t make any sense to start with a filter larger than the inputs, so my guess is that this topic in the lecture is where Prof Ng is explaining how the striding of the filter works. As you step the filter across the image (with padding or without), it will eventually step off the edge in either the horizontal or vertical direction. So you don’t get an output in those cases in which the filter is not completely contained. In these first few lectures explaining the fundamentals of convolutions, he also mentions why the most common practice is to use filters with odd numbers of pixels, e.g. 3, 5, 7, because with even sized filters, you can hit the case where some rows or columns of pixels at the edges of the image don’t get covered without stepping off the edge.
2 Likes