On Wed, Feb 10, 2010 at 8:38 AM, Tim Starling <tstarling(a)wikimedia.org> wrote: Just switching the thumbnailer should be sufficient, I agree the pile of code and the retries were fairly lame (and I think I complained about it on IRC). I'm not sure why any support for thumbnailing ogv's with ffmpeg was retained. I don't see how you can fix it in normalizeParams call unless you've scanned the stream and know where the keyframes are. Ffmpeg could be fixed, of course, but the ogg demuxer basically needs a rewrite... I think the patch you did to ffmpeg a while back was a lot better than the code they ultimately included. Here is a file that won't thumbnail under the current code: http://myrandomnode.dyndns.org:8080/~gmaxwell/theora/only_one_keyframe.ogv ffmpeg -y -ss 5 -i only_one_keyframe.ogv -f mjpeg -an -vframes 1 foo.jpeg throws a pile of errors then foo.jpeg is a zero byte file.

On Wed, Feb 10, 2010 at 8:17 PM, Tim Starling <tstarling(a)wikimedia.org> wrote: 256-frame keyframe interval is the default for anything libtheora 1.1 based in two-pass mode... So, yes. I'm pretty sure thats relevant. I wouldn't have even realized the existing processing chain was failing on files that contained no keyframes past the midpoint were it not for the fact that I'd bumped into failed thumbnailing on short files. ::shrugs:: There isn't really a reasonable "fixed heuristic distance between keyframes", the encoder can put them anywhere subject to the limitation that they not be more than KEYINT away from the last one. So even knowing the interval the locations can't be reliably determined without scanning the file. The alternative "one line fix" is switching to a thumbnailer that actually thumbnails the requested location. All the better when the thumbnailer correctly handles 4:2:2 and 4:4:4 colorspaces. Just saying——

On Wed, Feb 10, 2010 at 9:36 PM, Tim Starling <tstarling(a)wikimedia.org> wrote: So a couple months back mdale suggested using oggThumb, with the old installed ffmpeg was making spaghetti of some files (old ffmpeg not completely implementing the Theora spec) and the new one that (whomever) tried installing mage spaghetti in a different way (failing to thumb because ffmpeg didn't take your seeking patch eons ago). I'd never heard of it, went to look, and recoiled in horror. Then I sent a patch. Of the issues I raised, seeking was the only I didn't fix. Unfortunately oggvideotools reimplements libogg in C++ so it could use C++ memory management, my patience ran out before I got around to implementing it. If you search the archive you can see how strongly opposed I am to tools that linear scan unnecessarily. But 30x slower on a file that small sounds a bit odd. It's in SVN. After the author of the package applied my patches (on the same day I sent them) Mdale asked if he should delay Wikimedia deployment until the fixes I sent in went in, the author offered to simply do a new release. No one took him up on the author. I'm also not especially keen on its rather non-unixy style. Then again, I think C++ is pretty much crap too, so you can see what my opinion is worth. What I can say is that speaking from personal experience the author of this package is friendly, pleasant to work with, and responsive. Though 'submit patches' takes me out of the 'one line fix' I advertised, — sorry, I'd assumed that Mdale had already worked out the operational angles and my only concerns were correct output and not allowing it to be an enormous DOS vector. Cheers.