A Comparison study of the Local Feature Detectors

In the first part of this report, the performance of four state of art detectors is given, both in Repeatability Check and Image Retrieval. In the Repeatabiilty and Matching test,MSER works best in the viewpoint and illunimation change, while SIFT,SURF and Hessian Affine are proved to be robust to the scale changes, image Blur and JEPG compression.On the other hand,in the Image Retrieval Test,SURF performs best with respect to top recall accuracy, while Hessian Affine works best in the MAP. Next, the theoretical analysis of the three best derivatives based detectors is investigated thouroughly. The pipeline of the popular feature detection, including scale selection, region construction, and descriptor, is analyzed. In the scale selection, firstly, the criteria of a good selection function is systematically studied and outlined, secondly, under this criteria the ideal framework of the selection function is proposed, thirdly, the available algorithms are compared under this framework, and finally, a J function is proposed tentatively. In the stage of the region construction, firstly detained construction appoarches of each detector are discussed, and then the remaining problems of Hessian Affine Detector is pointed out for the future research. In the end, the advantages and disvantages of the most popular descriptors are discussed.