For the life of me, I can't find a suitable way to do this.
I have 3 videos (all h264), which supposedly came encoded from the same source (they have the same content / offset by up to a second). I do not have the original raw video / any reference video to use for comparison, so full reference and reduced metric tools do not apply here.
I have found 3 tools for no-reference metrics:
- http://vq.kt.agh.edu.pl/metrics.html
- https://github.com/Telecommunication-Telemedia-Assessment/bitstream_mode3_p1204_3
- http://compression.ru/video/quality_measure/vqmt_download.html
But none of these are suitable.
Unless I'm doing something seriously wrong, ffmpeg when converting these files to their raw, decoded yuv counterparts, the size is roughly 4.2 GB/min of content. The videos are all roughly 25 minutes long, so decoding a single one would mean > 100GB. The first tool, while seemingly what I want, requires raw footage. This means I'd need at least 300GB of free space to compare these videos, which while I guess is doable, I don't consider reasonable (nor would most people that don't work with video often).
The second tool appears to require a minimum quality of 1440p, not 1080p, and is of course only one specific metric. While I'd be okay with using only one metric at this point, I don't believe I can use this tool due to the whole 1080p vs 1440p mess.
The last tool is not open source and therefore completely unverifiable, and is not free for HD videos (this isn't something I'm willing to drop $300+ for a few-time-use thing, especially when the metrics within aren't really verifiable to be valid measures).
Is there any way (maybe with just ffmpeg?) to get no-reference metrics for 1080p, h264 video (and for that matter, h265 as well would be nice, but not necessary)? Unaware if this should be on the video forum, as I've seen previous related questions about comparison based metrics here on superuser.
Note that methods like VMAF and most VMQT tooling is full-reference, not no-reference, albeit plenty of academic literature being available for the latter. Full-reference does not fit my use case, because I do not have a reference video source. I have 3 (potentially lossy) outputs of different caliber. I know that the 3 sources are different, simply because even adjusting for offsets in frames, a simple video difference or xor filter show clear differences in the two videos (but I can't trust my eyes, nor monitor, to tell me which source is better qualitatively, and am willing to use computed metrics at this point).
