Determine whether FloWM’s mix layer can extend motion beyond its built-in velocities

Determine whether FloWM’s learned mix layer can extend the memory’s motion representation beyond the five built-in velocities and thereby carry blocks moving diagonally or at speed two without explicitly adding additional velocity-shift copies.

Background

FloWM represents motion using five built-in velocity copies: stationary motion and one-cell-per-step motion along each of the four cardinal directions. The varied setting includes diagonal motion and velocities of up to two cells per step, most of which fall outside this built-in set.

The learned mix layer can move content by one cell and might therefore, in principle, combine with a one-cell shift to represent two-cell motion. However, the experiments used only one training seed per run and produced weak results, so they do not establish whether the mix layer has this capability. Explicitly covering all 24 nonzero velocities would require 25 grid copies, substantially increasing the memory size.