Update some comments.

[nageru] / sobel.frag

diff --git a/sobel.frag b/sobel.frag
index 4fd12a37d40025241bd7bf8f5dab80eb63146fa2..9feab1ee6b26ba5c614fd5f0ad2cc920e31f58eb 100644 (file)
--- a/sobel.frag
+++ b/sobel.frag
@@ -11,12 +11,19 @@ void main()
        // There are two common Sobel filters, horizontal and vertical
        // (see e.g. Wikipedia, or the OpenCV documentation):
        //
-       //  [-1 0 1]     [ 1  2  1]
-       //  [-2 0 2]     [ 0  0  0]
-       //  [-1 0 1]     [-1 -2 -1]
+       //  [1 0 -1]     [-1 -2 -1]
+       //  [2 0 -2]     [ 0  0  0]
+       //  [1 0 -1]     [ 1  2  1]
        // Horizontal     Vertical
        //
-       // Computing both at once allows us to get away with eight
+       // Note that Wikipedia and OpenCV gives entirely opposite definitions
+       // with regards to sign! This appears to be an error in the OpenCV
+       // documentation, forgetting that for convolution, the filters must be
+       // flipped. We have to flip the vertical matrix again comparing to
+       // Wikipedia, though, since we have bottom-left origin (y = up)
+       // and they define y as pointing downwards.
+       //
+       // Computing both directions at once allows us to get away with eight
        // texture samples instead of twelve.
 
        float x_left   = tc.x - inv_image_size.x;