This was discussed in 2009 at VCEG in the context of AVC if my memory is correct. I don't recall there were exact numbers on the quality gain, at least not cross-checked ones. But it should be in the same range as the gain offered by the smoothing filter for intra prediction in HEVC. There is some, but probably not something to get too excited about.
--Dake -----Original Message----- From: Minhua Zhou [mailto:[email protected]] Sent: Tuesday, August 18, 2015 11:56 AM To: Dake He; Jean-Marc Valin; Timothy B. Terriberry; [email protected] Subject: RE: [video-codec] Questions about Thor Any idea about the quality again by using the de-blocked reference samples in intra prediction? Agree that it is possible technically, but architecture wise this would create a nightmare for HW. Of course, if there was significant gain, we could take a second look. -- Minhua -----Original Message----- From: video-codec [mailto:[email protected]] On Behalf Of Dake He Sent: Monday, August 17, 2015 1:01 PM To: Jean-Marc Valin; Timothy B. Terriberry; [email protected] Subject: Re: [video-codec] Questions about Thor Minhua might not like it, but still technically possible (otherwise he would not have asked the question in the first place) with careful ordering of processing steps of deblocking (vertical/horizontal) and prediction. --Dake -----Original Message----- From: video-codec [mailto:[email protected]] On Behalf Of Jean-Marc Valin Sent: Monday, August 17, 2015 3:50 PM To: Timothy B. Terriberry; [email protected] Subject: Re: [video-codec] Questions about Thor On 08/17/2015 03:33 PM, Timothy B. Terriberry wrote: > Well, certainly if you implement it in the order Jean-Marc was > describing, you could apply deblocking to pixels along interior edges > before using them to predict adjacent blocks along an exterior edge, > but I believe that was one of the "not implementation friendly" things > Minhua Zhou was complaining about. It means your hardware can't > pipeline reconstruction of those neighboring blocks until it has the > deblocked pixels, which is fairly late in the process. Well, most intra predictors use only the pixels on the edges of the block, so the interior pixels -- deblocked or not -- aren't used at all. In that case, the "recursive deblocking" is just equivalent to decoding the entire image and then deblocking rows/columns that are multiples of 8 but not 16 first, then deblocking those are are multiples of 16 but not 32, then deblocking the rest... Jean-Marc _______________________________________________ video-codec mailing list [email protected] https://www.ietf.org/mailman/listinfo/video-codec _______________________________________________ video-codec mailing list [email protected] https://www.ietf.org/mailman/listinfo/video-codec _______________________________________________ video-codec mailing list [email protected] https://www.ietf.org/mailman/listinfo/video-codec
