- The document describes an algorithm called seed-and-spawn that searches for matching local regions between two images by propagating an initial set of seed matches. It does this by spawning new matches from existing matches at different image scales and locations, continuously matching regions at coarser and finer scales.
- The algorithm takes an initial set of seed matches and refines them before adding them to a priority queue. It then repeatedly spawns new matches from the highest priority match and refines the spawns. Matches that satisfy criteria are added to the solution.
- The algorithm was evaluated on an object recognition benchmark and shown to significantly increase the discriminative power of a state-of-the-art local feature matching method.