tag:blogger.com,1999:blog-5246987755651065286.post3245748192065837829..comments2023-05-04T15:15:27.117-07:00Comments on cbloom rants: 10-22-10 - Some notes on Chroma Samplingcbloomhttp://www.blogger.com/profile/10714564834899413045noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-5246987755651065286.post-80442546409582409362010-10-26T23:02:06.951-07:002010-10-26T23:02:06.951-07:00Also, I should mention I actually *tried* this, us...Also, I should mention I actually *tried* this, using RMSE, and observed the ringing problem you reported.<br /><br />I tried augmenting the linear programming problem with monotinicity constraints:<br /><br />If x_0 < x_1 < x_2 < x_3, then require downsampled x_01 < x_23, but I don't recall ever getting any good results from this. Also monotinicity may not be good enough, maybe you want to make sure that if a gradient is smooth it stays smooth, not just monotonic; e.g. if the derivative (finite difference) over six samples is montonically increasing, than so should the derivative (finite difference) over the the "corresponding" downsampled ones. (Using sloppy pixel-is-a-little-square-box-filter terminology.)Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-5246987755651065286.post-55658504072262418602010-10-23T11:07:14.649-07:002010-10-23T11:07:14.649-07:00We know what the decoder will do to upsample, so i...<i>We know what the decoder will do to upsample, so instead we just take the idea that our encoder should output the coefficients which will make the best result after upsampling.</i><br /><br />I've long argued this as an approach to consider for mipmap generation: we know they're reconstructed with bilerp, so optimize the downsample of the mipmaps around reconstructing as best as possible after bilerp.<br /><br />Of course, we don't actually <i>display</i> texture mipmaps upsampled this way, so it's maybe not actually a good argument (instead they are better off with the bilerp resampling being a good approximation to a good-looking downsampling/blur of the source data, in which case we're back where we started).Anonymousnoreply@blogger.com