Anyway, the trick is, you take a low discrepancy sequence on a sphere (I used R[2] and a Lambert azimuthal projection) and use those angles as lines to sample depth – a measure of centrality – along; you map these angle samples over your points and reduce with min(), and you get an approximation of Tukey’s multivariate depth, which is a generalization of the median (or quantile). I may explain this better eventually but then again I may not.

And I should say: the way this actually works in practice is you take the timeseries of each pixel in the video, in oklab or your preferred color space, and do an interquartile (or similarly trimmed) mean on it.

@vruba yeah after our previous conversation about this a few months ago, I tried a bunch of things and it only got worse too. the Axis cameras are aggressively compressing but also doing a very destructive unsharp mask/sharpening thing first. hopeless.

@midendian Yeah, this example is cherry-picked for minimal camera shake, and I downsampled before doing anything, and I’m pretty sure you can still find shake-induced artifacts if you look.