Because when we interpolated during the upsampling process, the green and blue are "blended" together at the edge of the ball. So then later when we put it over the background, the blended green appears at the edge of the ball
atarng
Without premultiplication, the blue object is blended together with the green background (without knowing that the green background is supposed to be transparent), and the bits of green get left over when we apply the alpha on top of the color.
peanut
Agree, and I think this question is actually detailed explained in the next slide.
keenan
Yep! Good answers. :-)
bcagan
So fringing occurs because information from some unrelated background is accidentally being upscaled with the color information of the object?
Osoii
At first I didn't get it that how the upsampling is done by bilinear interpolation. Then I realize you can imagine the upsampling as a bigger square shape, and the original image is a texture, and you use bilinear interpolation to do the texturing, which is the upsampling.
barath
so, after some Algebra, I found that the difference between Pre-Multiplied and Non-pre Multiplied is a factor of 1/(alphaA + alphaB - alpha*alphaB). So, will multiplying Non-pre Multiplied answer by the above factor lead to a Pre-Multiplied Equivalent?
xiaol3
The edge would be blended with blue and green if we don't premultiply.
diegom
Yeah echoing what was said above, the bilinear interpolation used for upsampling blended a bit of green into the fringe of the blue blob there. I'm curious as to what would've happened if we hadn't upsampled. In theory, we could still blend the original images that aren't upsampled and maybe those wouldn't have this fringe artifact?
Azure
Yeah, I think the issue that's causing the green fringe is the interpolation during upsampling. By premultiplying we kinda get rid of this, causing the fringes to interpolate with a black background which is less conspicuous/fades to "nothing"/combines well with background.
Because when we interpolated during the upsampling process, the green and blue are "blended" together at the edge of the ball. So then later when we put it over the background, the blended green appears at the edge of the ball
Without premultiplication, the blue object is blended together with the green background (without knowing that the green background is supposed to be transparent), and the bits of green get left over when we apply the alpha on top of the color.
Agree, and I think this question is actually detailed explained in the next slide.
Yep! Good answers. :-)
So fringing occurs because information from some unrelated background is accidentally being upscaled with the color information of the object?
At first I didn't get it that how the upsampling is done by bilinear interpolation. Then I realize you can imagine the upsampling as a bigger square shape, and the original image is a texture, and you use bilinear interpolation to do the texturing, which is the upsampling.
so, after some Algebra, I found that the difference between Pre-Multiplied and Non-pre Multiplied is a factor of 1/(alphaA + alphaB - alpha*alphaB). So, will multiplying Non-pre Multiplied answer by the above factor lead to a Pre-Multiplied Equivalent?
The edge would be blended with blue and green if we don't premultiply.
Yeah echoing what was said above, the bilinear interpolation used for upsampling blended a bit of green into the fringe of the blue blob there. I'm curious as to what would've happened if we hadn't upsampled. In theory, we could still blend the original images that aren't upsampled and maybe those wouldn't have this fringe artifact?
Yeah, I think the issue that's causing the green fringe is the interpolation during upsampling. By premultiplying we kinda get rid of this, causing the fringes to interpolate with a black background which is less conspicuous/fades to "nothing"/combines well with background.