On Tue, Jul 27, 2010 at 4:55 PM, Brad Kemper <brad.kemper@gmail.com> wrote: > If you have a picture of what it is supposed to look like, with > the predicted values rendered, then you could more easily make an initial > visual comparison. Sure, there's nothing *wrong* with making an extra picture and doing pixel comparisons that way. It's just not necessary for the actual conformance testing. > But the approximation would fail your test, because you're comparing each > pixel's value against the value of an actual gaussian distribution. Firefox > would fail, because it has noticeable banding, whereas an actual gaussian > distribution presumably wouldn't. I think you might be confused - Firefox passes my suggested conformance criteria just fine (if we pretend that it treats the blur length properly). It is, at worst, 6 color units away from a true gaussian, which is within the 5% boundary. The conformance criteria is that each pixel must be within some percentage of a true gaussian. Banding like you see in Firefox near the edges is perfectly acceptable as long as it remains within the specified percentage of the reference value. ~TJReceived on Wednesday, 28 July 2010 00:13:22 UTC
This archive was generated by hypermail 2.4.0 : Monday, 23 January 2023 02:13:48 UTC