. What is the key difference between true convolution and correlation?
Convolution flips the kernel before sliding; for symmetric kernels this makes no difference, but for asymmetric kernels (like Sobel) it does.
. Does cv2.filter2D compute convolution or correlation?
Most image libraries, including OpenCV's cv2.filter2D, actually implement correlation despite being casually called "convolution".
. What does it mean for a 2D kernel to be separable?
A separable kernel like the Gaussian can be applied as a 1D pass along x then a 1D pass along y, which is much cheaper than the full 2D kernel.
. In the Sobel edge-detection example, why does the convolution-vs-correlation distinction actually matter?
For symmetric kernels (box, Gaussian) flipping changes nothing, but asymmetric kernels like Sobel give different results under convolution vs. correlation.