One part of a window-based object detector is scaling the input to different resolutions before running the window across it (one can also scale the window, but this is not efficient on modern cpus).
So i've been looking at up/down resampling in little bits here and there over the last few days.
To cut a long story short, after coming up with a fairly complex but pretty good (i think) implementation of a one-pass 1/N and 2/N 2-D 'up-sample/down-sample' scaling filter ... I found that using a simple 1/N box filter is more than good enough for the application - and about 2x faster. Should NEONise pretty easily too.
The up/down filter may well be useful for other purposes though and I did learn more about up/down filters in general which is something i've been meaning to do. Maybe at some point i'll write about both.
I was only looking at implementing this for ARM but the algorithm I came up with should fit epiphany quite well - so at some point I will look deeper into it. Epiphany does offer other more exotic options mind you.