a few days ago naoyuki uchida put out the oklch paper, a clean three parameter extension of oklab, and it raised a fair point that my helmlab metricspace looks like it overfits. instead of just replying i said i would come back with a proper data backed answer, so i ran the test myself on every applicable dataset with one stress definition for everyone. here is the whole thing as a short q and a thread
01
q: does helmlab metricspace overfit
a: yes but its source dependent, not global. the oklch paper by naoyuki uchida raised this about Helmlab Metricspace so i sat down and stress tested it properly. i put metricspace up against oklch , ciede2000, cielab, oklab and cam16 ucs on 8 color difference and discrimination datasets, all scored with the exact same garcia stress so its apples to apples. quick map of the result below, the rest of the thread breaks it down