An in-depth study of local image descriptors and their performance

 Master Defense by Anders Boesen Lindbo Larsen


 In this project, we study the performance of local image descriptors using a recently released dataset [2] that offers unique evaluation possibilities compared to previous datasets. Our investigation is twofold in that we evaluate current state-of-the-art descriptors and develop our own descriptors to explore different designs and their performance implications.

The descriptors we develop are based on the locally orderless image representation and on higher-order differential structure. We show that we by the use of higher-order image structure are able to reduce the descriptor dimensionality and partly do away with the multi-local description, that has previously been considered crucial for achieving good performance. We test the descriptors in a variety of scenarios (namely, large changes in scale, viewing angle and lighting) and find that the choice of descriptor should be based with the image data and the end application in mind.

Supervisor: Kim Steenstrup Pedersen

Censor: Rasmus Larsen, IMM DTU