2D Convolution Explained: Fundamental Operation in Computer Vision

[@LearnOpenCV]

Check out our FREE Courses at OpenCV University : opencv.org/university/free-courses/

[@DillanWilson-q6b]

Great video. I have a small nitpick: animation at 0:05 is showing a cross correlation not convolution. For convolution, you flip the kernel over each axis before performing the dot product

[@simonhofmann8387]

Yes, but I think he did that without talking about it. Otherwise his Sobel Kernel would be incorrect (it is already flipped). I was wondering too :D

[@usmankhawar445]

I have one confusion, I have been searching for this but couldn't find any concrete answer. Is it correct to have the x-derivative and y-derivatives having dimensions smaller than image matrix? shouldn't these two have the same dimensions as well. Any one could guide me please.

[@HuCEcpvrLab]

Should you not apply the filter mask values mirrored in the dot product? Unmirrored the operation is a correlation. Especially if the filter is an asymmetric filter such as the sobel is.

[@LearnOpenCV]

Hi, you are right. For asymmetric filters like Sobel, true convolution involves mirroring the filter, distinguishing it from correlation. In many ML frameworks, the distinction is blurred as filters are learned, not predefined. However, for correct edge orientations with Sobel, we must apply the mirrored filter.

[@kkuzmicki99]

Great video

[@matteo2608]

Hi I have a question, Is it right to flip the kernel and then do the convolution? Thank you

[@LearnOpenCV]

You can do that according to your application needs. Flipping the kernel will just create another kernel, anyway.

[@jb14_99]

Thanks for making the video!

[@LearnOpenCV]

Thank you very much!

[@dinazul233]

good one!

[@LearnOpenCV]

Thank you!