Ives Rey Otero (North Carolina State University)
Thursday, 27.10.2016, 11:30
The SIFT method is the first efficient image comparison model. It was the first method to propose a practical scale-space sampling and to put in practice the theoretical scale invariance in scale-space theory. SIFT associates with each image a list of translation, rotation and scale invariant features used for comparison with other images.
Despite countless applications and an avalanche of variants, not much has been done to really understand this central algorithm and to find out exactly what improvements we can achieve.
The work I will be presenting provides a meticulous dissection of SIFT's complex chain of transformations, from the extraction of invariant keypoints to the computation of feature vectors. It discusses the exact computation of the Gaussian scale-space, at the heart of SIFT as well as most of its competitors. It defines a rigorous simulation framework to find out how the stability of keypoints is influenced by the sampling of the scale-space sampling parameters, and other perturbations such as image blur, aliasing and noise.
This in-depth analysis shows that, despite numerous methods claiming to outperform SIFT, there is in fact limited room for improvement in methods that extract keypoints from a scale-space. It also shows that the performance analysis of local feature detectors, mainly based on the repeatability criterion, is biased towards methods producing redundant (overlapping) descriptors. By using an amended evaluation metric, SIFT is shown to outperform most of its more recent competitors.
2015 - 2016: North Carolina State University, Postdoctoral Researcher
2010 - 2015: ENS Cachan, PhD in Applied Mathematics, under the supervision of Jean-Michel Morel and Mauricio Delbracio
2009 - 2010: ENS Cachan, MSc in Applied Mathematics, Master "Mathematics, Vision, Learning"