How much faster do we expect this approximation to be over just applying the 1-pixel box filter?
Staffi-geng
Here, the “approximation” means that we aren’t geometrically/analytically calculating the exact average value of a pixel. Convolution by a 1-pixel-width box filter whose kernel dimensions are NxN is equivalent to NxN supersampling!
colinsteidtmann
I think it should be made more clear that we'll need to create images at high resolutions and then sample from those. From the lecture, it seems like we would be able to sample a single pixel NxN times, but I don't see how that's possible as a pixel is the smallest unit (I think)
spegeerino
If we were working with just trying to rasterize abstract shapes (maybe a 2D game graphics engine), would it be possible to define a triangle by giving a distance function which tells you how far away you are from the border of the triangle, and then color the pixels on a gradient by depending on how far the pixel is from the border? That way it's much less computationally expensive (you can essentially compute a good approximation of the supersampling average without actually having to sample n^2 times as many points)
angelajyzhang
I wonder at what point does the level of supersampling not trade off for the computational work a program will do? i.e. if we do 16x16 supersampling, would the quality be so good that it's worth running supersampling at that level?
How much faster do we expect this approximation to be over just applying the 1-pixel box filter?
Here, the “approximation” means that we aren’t geometrically/analytically calculating the exact average value of a pixel. Convolution by a 1-pixel-width box filter whose kernel dimensions are NxN is equivalent to NxN supersampling!
I think it should be made more clear that we'll need to create images at high resolutions and then sample from those. From the lecture, it seems like we would be able to sample a single pixel NxN times, but I don't see how that's possible as a pixel is the smallest unit (I think)
If we were working with just trying to rasterize abstract shapes (maybe a 2D game graphics engine), would it be possible to define a triangle by giving a distance function which tells you how far away you are from the border of the triangle, and then color the pixels on a gradient by depending on how far the pixel is from the border? That way it's much less computationally expensive (you can essentially compute a good approximation of the supersampling average without actually having to sample n^2 times as many points)
I wonder at what point does the level of supersampling not trade off for the computational work a program will do? i.e. if we do 16x16 supersampling, would the quality be so good that it's worth running supersampling at that level?