ICCV Daily 2019 - Thursday

4 Poster Presentation DA I L Y Kwang Moo Yi is an assistant professor at the University of Victoria in Canada, having made the leap from developer to faculty. He was previously a postdoc researcher at École Polytechnique Fédérale de Lausanne (EPFL). He speaks to us following his poster on Tuesday. The work is about learning local features. Trying to figure out where you will be able to align two images together or find the camera pose difference between images. The state- of-the-art way to do this is by using SIFT features developed 20 years ago by David Lowe . This work has tried to do a learned counterpart for that. It’s all about the last part of the pipeline which is to describe the points that are selected in a robust way. While at EPFL, Kwang Moo worked on learning local features, resulting in the LIFT paper. A downside they found was that the detectors were making a lot of mistakes in the scale and orientation estimation part , meaning how you would look at these local patches. What does this work do differently? He explains: “This work is trying to use a different representation for looking at these local patches and to have them so that they would be represented in something called the log-polar representation which turns images into something looking like circles.Whenyou rotate or scale the images it becomes translation in this representation, and we all know deep networks are very good at learning to compensate for translational errors.” Kwang Moo tells us the most challenging part is getting the benchmark right. A lot of the baselines have their own default settings which might not be the best for a given dataset. They had to do a lot of under- the-hood work to sort that out which took a long time. From an algorithmic point of view, he says solving the Beyond Cartesian Representations for Local Descriptors "The most challenging part is getting the benchmark right."