tag:blogger.com,1999:blog-5246987755651065286.post7381815863773719145..comments2024-02-22T16:15:42.388-08:00Comments on cbloom rants: 05-26-09 - Some Study of DCT coefficientscbloomhttp://www.blogger.com/profile/10714564834899413045noreply@blogger.comBlogger7125tag:blogger.com,1999:blog-5246987755651065286.post-39615228762687141072009-05-29T18:37:45.318-07:002009-05-29T18:37:45.318-07:00"aren't things like the number of zeroes going to ..."aren't things like the number of zeroes going to depend heavily on the choice of quantizations? which are you doing here?"<br /><br />I'm always doing uniform quantization here BTW.<br /><br />Changing quantizer absolutely changes the # of zeros, but it shouldn't affect whether a value being zero is an independent event or not.<br /><br />My comment about the typical cases is based on my belief that this type of coder should only be used in the 0.25 - 1.0 bpp range.<br /><br />"also, on the PCA blog post, the same issue crossed my mind: if you found the optimal transformation to some arbitrary basis, how would you quantize in that basis?"<br /><br />Well you're giving up any attempts at perceptual quantization at that point. Uniform quantization in any orthonormal basis is equivalent in an R-D sense.<br /><br />There is the issue that when you quantize you wind up seeing the quantization shapes. How that turns out exactly in human perceptual error is hard to say (though personally I find the claims that DCT are good in this sense to be quite dubious).cbloomhttps://www.blogger.com/profile/10714564834899413045noreply@blogger.comtag:blogger.com,1999:blog-5246987755651065286.post-33231211560452186082009-05-29T17:53:07.099-07:002009-05-29T17:53:07.099-07:00aren't things like the number of zeroes going to d...aren't things like the number of zeroes going to depend heavily on the choice of quantizations? which are you doing here?<br /><br />also, on the PCA blog post, the same issue crossed my mind: if you found the optimal transformation to some arbitrary basis, how would you quantize in that basis?Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-5246987755651065286.post-22313859404592017412009-05-27T13:36:45.574-07:002009-05-27T13:36:45.574-07:00" I assume the grayscale values are in linear s..." I assume the grayscale values are in linear scale, i.e. 50% gray = a correlation with magnitude of 0.5?"<br /><br />It's linear scale and actually gamma corrected, so apparent 50% brightness = correlation 0.5 (not pixel value 128).<br /><br />Yeah, I agree it's somewhat shocking when you see it visually just how much correlation there is. These results are all pretty well known, but it clarifies things mentally to see the pictures.cbloomhttps://www.blogger.com/profile/10714564834899413045noreply@blogger.comtag:blogger.com,1999:blog-5246987755651065286.post-49793181864043635002009-05-27T12:48:46.959-07:002009-05-27T12:48:46.959-07:00I assume the grayscale values are in linear scale,...I assume the grayscale values are in linear scale, i.e. 50% gray = a correlation with magnitude of 0.5?<br /><br />In any case, this is really interesting. The DCT is really obviously very far off from the optimal decorrelating transform for these images. It'd be really interesting to see how these images look for prediction residuals when doing H264-style coding.ryghttps://www.blogger.com/profile/03031635656201499907noreply@blogger.comtag:blogger.com,1999:blog-5246987755651065286.post-1737404905051908742009-05-27T11:06:36.055-07:002009-05-27T11:06:36.055-07:00"or it just intrinsically behaves differently from..."or it just intrinsically behaves differently from the AC coefficients?"<br /><br />this. <br /><br />For one thing you'd have to subtract off the global mean of the whole image or it would be scaled up by a huge bias.<br /><br />If you delta DC from the neighboring DC's it acts more like the other coefficients, but then we're getting into details that I didn't want to involve in this test.cbloomhttps://www.blogger.com/profile/10714564834899413045noreply@blogger.comtag:blogger.com,1999:blog-5246987755651065286.post-15976570685863414532009-05-27T10:30:14.232-07:002009-05-27T10:30:14.232-07:00How exactly is DC weird? Just the way JPEG encodes...How exactly is DC weird? Just the way JPEG encodes it (DPCM?) , or it just intrinsically behaves differently from the AC coefficients?won3dhttps://www.blogger.com/profile/09787472194187459747noreply@blogger.comtag:blogger.com,1999:blog-5246987755651065286.post-72358100223003479532009-05-27T09:48:58.478-07:002009-05-27T09:48:58.478-07:00I will win that argument by inexorable momentum. ...I will win that argument by inexorable momentum. Victory will be mine!Dave Moorehttps://www.blogger.com/profile/13438224000659728514noreply@blogger.com