A general framework for image feature matching without geometric constraints

• We propose a general framework of 4 algorithms for image feature matching.
• We prove that the traditional Ratio-Match is the worst performer.
• We confirm the theoretical results experimentally on over 3000 image pairs.
• 20 percentage-point higher precision without increased computation time.

Computer vision applications that involve the matching of local image features frequently use Ratio-Match as introduced by Lowe and others, but is this really the optimal approach? We formalize the theoretical foundation of Ratio-Match and propose a general framework encompassing Ratio-Match and three other matching methods. Using this framework, we establish a theoretical performance ranking in terms of precision and recall, proving that all three methods consistently outperform or equal Ratio-Match. We confirm the theoretical results experimentally on over 3000 image pairs and show that matching precision can be increased by up to 20 percentage-points without further assumptions about the images we are using. These gains are achieved by making only a few key changes of the Ratio-Match algorithm that do not affect computation times.