- From: Harald Alvestrand <harald@alvestrand.no>
- Date: Thu, 14 Aug 2014 18:21:15 +0200
- To: "Gili T." <cowwoc@bbs.darktech.org>, public-media-capture@w3.org
- Message-ID: <53ECE1FB.5060406@alvestrand.no>
On 08/14/2014 05:45 PM, Gili T. wrote: > > Hi Harald, > > Can you please explain how an epsilon of 1/1000 equates to one pixel > in HD? > The aspect ratio of a 1024x768 image is 1.3333333333 The aspect ratio of a 1025x768 image is 1.3346354166 The difference is 0.0013020833, which is quite close to 1/1000. > What able future proofing? What happens when we want 4k resolution in > the future? > On 4K resolution, the difference will be ~4 pixels. Most coding schemes have 16-pixel macroblocks. The point is - 1/1000 is *precise enough for all practical purposes*. Nobody NEEDS to specify an aspect ratio of 1.3334 and not have it match 1.3333. > Thanks, > Gili > > On Aug 13, 2014 1:11 PM, "Harald Alvestrand" <harald@alvestrand.no > <mailto:harald@alvestrand.no>> wrote: > > On 08/13/2014 12:11 PM, Gili T. wrote: >> >> And I'll repeat that different ratios will need different >> epsilons. Epsilons imply inaccuracy and there is no one number >> that will work for every use case. Why not just go with >> "numerator/denominator", parse it to two integers and compare >> with 100% accuracy? >> > > Did you intend for this to go to me only? > > 1/1000 is one pixel in HD. We don't need more precision. > > >> Gili >> >> On Aug 13, 2014 8:50 AM, "Harald Alvestrand" >> <harald@alvestrand.no <mailto:harald@alvestrand.no>> wrote: >> >> On 08/12/2014 06:17 PM, Jan-Ivar Bruaroey wrote: >> >> On 8/11/14 6:17 PM, Martin Thomson wrote: >> >> On 11 August 2014 14:34, Jan-Ivar Bruaroey >> <jib@mozilla.com <mailto:jib@mozilla.com>> wrote: >> >> Actually, the spec already says: "The exact >> aspect ratio (width in pixels >> divided by height in pixels), represented as a >> double rounded to the tenth >> decimal place" [1] >> >> So we effectively have our epsilon already: >> .0000000001 >> >> So no action required it seems. >> >> Oh, that's good, there's an epsilon; but it's bad. >> 1.777777778 isn't >> the same as 16/9 based on that. Nor is 1.7777777777. >> Add a single >> digit to either and it would match. I may have >> counted wrong. I'm >> sure that others will too. >> >> >> Feel free to propose a different epsilon I suppose. >> >> I presume this would have nothing to do with inaccuracies >> inherent in floating-point math then (or we could have >> picked an epsilon much closer to everyone's worst machine >> epsilon), but instead from a desire to accommodate people >> handwriting rounded decimal numbers for aspect. Just so >> we're clear on the properties we seek. >> >> >> I'll repeat my suggestion of an epsilon of 1/1000. >> >> >> .: Jan-Ivar :. >> >> >> >> >
Received on Thursday, 14 August 2014 16:21:48 UTC