At recognition time, we choose a frame of the second structure (the scene) to be the basis for computing the invariant representation and retrieve, thanks to the hash table, what are the compatible model frame couples. If the basis belongs to a common substructure, then a significant number of frames are in the same configuration with respect to it. We then match the model and scene bases.

This process is repeated for every amino acid as the basis to find its possible matches in the model and we only keep matches that score above a given threshold (typically 4 or 5 in our applications).