I have a vague memory of a series of papers working to reduce the constant factor in the number of comparisons for deterministic linear time median finding, using increasingly elaborate (but interesting!) techniques to do so.
In particular, I remember one of the papers using the term "green forests" for a data structure storing information from previous comparisons (in an attempt to reuse them and avoid future comparisons).
Does this ring any bells? "median finding green forests" turns out to be a terrible search term.