I worked on video encoding for a long time. And chroma-subsampling is no longer necessary.<p>In short, chroma subsampling was performed because it matched the data more closely to how the eyes work, at the cost of losing certain colors.<p>Today, modern video and image compression algorithms can compensate without using chroma subsampling. Starting with the full image actually gives the encoder more to work with, and it can optimize better.<p>Because the software has improved (at least, the cutting edge open source tools have), chroma subsampling is obsolete, and will produce inferior picture qualtiy at larger byte sizes than their rgb444 compressed alternatives.<p>If you are seeing improvements by doing chroma subsampling, either you are hurting quality or you should use better compression software.
Here's a worst case test image for chroma subsampling:<p><a href="https://commons.wikimedia.org/wiki/File%3ARicardo_Quaresma_%28L%29%2C_Pablo_Zabaleta_%28R%29_%E2%80%93_Portugal_vs._Argentina%2C_9th_February_2011.jpg" rel="nofollow">https://commons.wikimedia.org/wiki/File%3ARicardo_Quaresma_%...</a><p>I first saw this in Fabrice Bellard's BPG demonstration at <a href="http://xooyoozoo.github.io/yolo-octo-bugfixes/" rel="nofollow">http://xooyoozoo.github.io/yolo-octo-bugfixes/</a> ("Soccer Players"). Every lossy format fails to reproduce the red dots on the green stripe because they all use chroma subsampling. Until I see tools that can automatically deal with images like this I think the anti-subsampling attitude mentioned in the article is correct. "Unless you are willing to evaluate each image, the thinking goes, you should avoid subsampling."