Cool work! Do you think there’s any architectural bias that prevents learning extrema rather than only minima or only maxima? Or is it mostly the repeatability issue? Just has me thinking about classical SIFT DoG and Harris handling both light/dark

This is a fun example with a continuous transition between distinct 3D scenes!

Here's a reconstruction of a movie establishing shot

I’m not too familiar with it, but it seems there is some equivalence noted in Section 3.2 of “Feature preserving point set surfaces based on non‐linear kernel regression”.